Complete source rendering · 19 November 2026

O2AT v0.2 engineering draft.

This semantic HTML edition is generated from the bundled editable ODT. It retains source text, all physical tables, 28 footnotes, linked references and the embedded architecture SVG. Pagination and page styling follow the matching PDF.

← Specification overview · Download the matching PDF · Download the editable ODT

Contents · 49 section anchors

OPEN AUDIO ARCHITECTURE SPECIFICATION • VERSION 0.2

Open Optical Audio Transport (O2AT)

An open Layer-2 optical successor to TOSLINK, centred on low-power 1G SFP

This table is wider than the reading column. Scroll horizontally to read all columns.

o2at table 1
Specification fieldDetails
Protocol identifierO2AT in prose; o2at in code, identifiers and schemas
Version and statusv0.2 engineering draft for manufacturer and community review; proposed wire-version nibble 0x2
AuthorTunji Durodola (LinuxProphet)
Principal Systems Architect, pocketOne OÜ
Revision date19 November 2026
Mandatory physical baseline1000BASE-X at 1.25 GBd; 1000BASE-SX optics, 850 nm, duplex LC glass fibre; 1.000 Gb/s MAC service
Transport envelopePCM in 32-bit containers through 2,822.4 and 3,072 kHz; native packed DSD through DSD2048; 2–32 channels subject to admission
Ethernet framingRaw IEEE 802.3 Layer 2; local experimental EtherType 0x88B5; fixed 12-byte o2at header; standard 1,500-byte MTU
Clock architectureReceiver-owned conversion clock with elastic buffering and source feedback; external 10 MHz input mandatory for the reference tier
Optional expansion10GBASE-R for aggregation or channel counts exceeding the admitted 1G budget; not required for the stereo replacement profile
Open implementation policyProposed MIT software/RTL and CERN-OHL-P v2 hardware releases; openly published specification and interoperable test vectors

Revision focus

Version 0.2 makes 1G SFP the default endpoint design, doubles the earlier maximum format rates and replaces unsupported claims about optical modules with explicit budgets. It retains standard Ethernet framing, adds a precise receiver-paced control model and separates transport capability from existing DAC, USB and I²S limits. This document is a proposal, not a ratified standard or a claim that compliant products already exist.

1. Executive Summary & Problem Statement

O2AT proposes a durable replacement for the consumer optical audio socket: inexpensive duplex glass fibre, a standard pluggable optical module and an openly specified audio protocol. The primary connection is a direct source-to-DAC link that requires neither an IP address nor a network administrator. The core standard is 1G SFP; the optional 10G fabric is not part of the minimum consumer replacement requirement. Its purpose is to preserve TOSLINK’s most valuable property, a non-conductive signal path, while removing the installed base’s restrictive format envelope and dependence on an embedded audio clock.

The engineering centre of gravity in v0.2 is 1G SFP. A 1.25 GBd 1000BASE-X link offers a 1.000 Gb/s MAC service after line coding. With the framing defined here, stereo PCM32 at 3,072 kHz occupies 203.981 Mb/s including a VLAN tag, Ethernet framing, preamble and interpacket gap. Stereo native DSD2048 in the explicit 48 kHz-derived extension occupies the same rate. Eight channels occupy 815.923 Mb/s. These are calculated transport budgets, not announcements of commercial playback support.

The audio samples cross an asynchronous FIFO before conversion. The DAC-side clock sets the consumption rate; packet arrival determines only when data become available. Every source obeys receiver credits; coherent sample-clock discipline may additionally control its long-term production rate. This removes packet-arrival timing from the intended conversion-clock path. It does not remove the receiver’s own clock phase noise, supply coupling, metastability risks or the consequences of a FIFO underflow.

A reference-tier receiver shall provide a specified 10 MHz reference input, an independently engineered audio clock generator, lock monitoring and controlled holdover. Entry-level O2AT remains useful with an internal oscillator and receiver feedback. Requiring an external reference for every consumer installation would obstruct the central objective of replacing TOSLINK with a simple, dependable optical connection.

1.1 What changes in v0.2

  • The baseline is 1000BASE-X and 1000BASE-SX, with 10G retained as an optional fabric and high-channel-count profile. A mechanical SFP+ cage may host a qualified 1G module when its host electronics support that rate.
  • The format namespace extends PCM to 2,822.4/3,072 kHz and raw DSD to 90.3168 MHz per channel, plus an explicitly labelled 98.304 MHz extension. Exact integer rates replace the overflowing one-byte multiplier field in the attached draft.
  • The audio header is consistently 12 bytes. PCM is big-endian, left-aligned in 32-bit containers. DSD uses chronological, most-significant-bit-first blocks. Wire version 0x2 intentionally supersedes the incompatible draft header.
  • Ethernet PAUSE is treated as link-local congestion control. End-to-end receiver credit and upstream clock feedback address sustained rate mismatch. Generic switches do not confer a deterministic latency guarantee.
  • Manufacturer comparisons distinguish documented behaviour from proprietary implementation details. MSB Pro ISL, HiFi Rose optical USB and standard Ethernet optics share components without sharing a wire protocol.

1.2 Scope, terminology and conformance

“Shall” identifies a proposed v0.2 conformance requirement; “should” identifies a recommendation; “may” identifies an option. The normalised machine-readable prefix is o2at. A device shall advertise only formats it can accept, buffer and deliver at its physical audio output. A link can carry a format that the attached conversion engine cannot reproduce, so negotiation shall reject unsupported modes before audio starts.

O2AT is a Layer-2 protocol within a single bridged Ethernet domain. Switches can bridge its Ethernet frames; ordinary IP routers do not route them. The prototype EtherType is experimental and shared with other experiments, so v0.2 is appropriate for controlled evaluation networks pending a production identifier and a completed interoperability programme. The proposal neither claims universal patent clearance nor grants rights over third-party standards or vendor IP cores.

4. O2AT Design Objectives & Architectural Principles

The replacement must be understandable to a listener and precise enough for two independent engineering teams to implement. A direct optical connection should discover its peer, agree a supported format, prefill a bounded buffer and start safely. It should not require the listener to choose clock-recovery constants, FPGA pins or a network addressing scheme.

4.1 Three implementable profiles

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 3. Profile requirements
ProfileRequired behaviourPermitted expansion
o2at-basic1G optical Ethernet, direct-link discovery, stereo PCM32 containers at 44.1 and 48 kHz with 16/24/32 valid bits, capability negotiation, receiver credits, safe mute and diagnosticsAdditional PCM rates, native DSD and more channels when explicitly advertised
o2at-referenceAll basic functions plus a documented 50 Ω 10 MHz input, selectable internal/external clocking, lock/holdover reporting and measured clock performanceHigh-rate PCM/DSD and multiple channels as validated; external reference is not required to obtain link-up
o2at-fabricStandard Layer-2 bridging and admission across documented switch configurations; endpoint profiles remain explicitOptional 10GBASE-R uplinks and endpoints for aggregation or rates beyond 1G admission

A high-rate capability claim is additional to the profile name. A manufacturer may advertise “PCM32/3072, DSD2048/44.1, 8 channels at 1G” only after validating those combinations. The high-rate envelope does not force every affordable TOSLINK replacement to implement a multi-megahertz conversion engine.

4.2 Isolation, openness and a stable channel model

The fibre assembly shall contain no conductive path between the two O2AT ports. This isolates the signal link galvanically. It does not guarantee isolation of the entire installation when protective earth, USB, analogue interconnects or shared coaxial reference-clock distribution also connect chassis. Safety-earth arrangements must remain governed by the equipment design; the protocol cannot cancel an external conductive path.

PCM and DSD shall be transported without implicit volume control, sample-rate conversion, channel reordering or DoP insertion. Any processing belongs to an explicitly selected endpoint function outside the transparent transport mode. Two-channel ordering is left then right. Additional channels use a negotiated map with defined logical channel identifiers, replacing the undocumented clock/data/flag assignments associated with I²S over HDMI.

Reference RTL, packet encoders, dissectors, schematics and conformance vectors should be published under clear open licences. The proposed release policy is MIT for software/RTL and CERN-OHL-P v2 for hardware, with a separately stated open document licence for the specification. Third-party PHY IP and vendor tools retain their own licence conditions. No existing open reference implementation is claimed by this paper.

5. Physical Layer & Wire Protocol Specification

5.1 Mandatory 1G optical baseline

The baseline shall implement an IEEE 802.3 Ethernet MAC with 1000BASE-X PCS/PMA at 1.25 GBd, full duplex. Its common optical profile is 1000BASE-SX at 850 nm using matched duplex-LC modules and OM3 or OM4 multimode glass fibre. Optional 1000BASE-LX single-mode links shall be identified explicitly and use qualified optics and fibre at both ends. A connector alone does not identify wavelength, launch power, receiver sensitivity or Ethernet compatibility.16

The source and receiver each contain a transmitter and a receiver. One fibre carries source-to-sink audio; the other carries discovery, control, credits and status in the opposite direction. Ethernet capacity is full duplex, so reverse control traffic does not consume the forward link’s serial bandwidth. MAC processing and buffers still require capacity for both directions.

A host shall implement module presence, transmit disable, loss-of-signal and fault handling, appropriate power sequencing and the required electrical interface. It should expose module identity and supported diagnostics. Only qualified module classes should be enabled. A 10G-only host cannot be presumed to operate a 1G SFP, and replacing an optic does not change the host PCS or SerDes rate. The baseline shall implement and enable 1000BASE-X auto-negotiation, advertising full duplex. Forced operation is an explicit paired-test option. Optional PAUSE capability shall be advertised and configured consistently; speed and duplex shall be reported.

The receiver reference design should begin with a low-power optical module, preferably a qualified part below 0.5 W, but this is an endpoint engineering target rather than a redefinition of an MSA power class. Coherent’s FTLF8519P3xyL provides a documented example below 500 mW. The complete budget must add the cage interface, FPGA transceiver, PCS/MAC, memories, regulators and clock circuitry.17

5.2 Revised SFP and SFP+ engineering assessment

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 5. A qualified engineering comparison, replacing universal power and CDR claims
Parameter1G SFP baseline10G SFP+ optionDesign conclusion
Serial rate and MAC capacity1.25 GBd; 8b/10b; 1.000 Gb/s MAC service10.3125 GBd; 64b/66b; 10.000 Gb/s MAC serviceNeither MAC figure is net audio capacity. Include all packet overhead and admission margin.
Optical module powerNamed example: Coherent FTLF8519P3xyL below 500 mWNamed example: Cisco 10GBASE-SR variants specify 1 W maximum1G is a sensible low-power starting point. Do not impose a universal 2× or 3× power ratio.
CDR / internal retimingMany 1G modules rely on host clock recoveryAn internal CDR is not mandatory for every SFP+ moduleSFF-8472 identifies optional retimer/CDR presence. Qualify the actual module and host pair.
RF and supply couplingLower serial rate generally eases implementationHigher rate can increase SerDes and power-integrity demandsMeasure coupled noise. Baud rate is not a single EMI fundamental or evidence of audibility.
FPGA interfaceCompliant 1.25 GBd SerDes or qualified external PHYCompliant high-speed SerDes and SFI channelOrdinary LVDS pins alone do not guarantee CDR, comma alignment or PCS compliance.
PCB material and layoutControlled impedance, return paths and validated losses remain necessaryShort standard-FR4 channels can be feasible within the applicable loss budget10G does not universally mandate an exotic laminate; qualify the actual stack-up and route.
Role in O2ATMandatory endpoint baseline; stereo and admitted eight-channel extreme formatsOptional 16/32-channel extreme formats, aggregation and shorter serialisationChoose 1G for the consumer socket. Preserve 10G where the payload or topology benefits.

Evidence: module power examples, optional CDR indication and host-channel guidance.18

5.3 Ethernet envelope and identification

8b/10b constrains run lengths and bounds running disparity. It does not produce an exact 50% optical duty cycle in every symbol. Its purpose includes reliable serial recovery and AC-coupled channel behaviour; a laser is not physically unable to remain on or off for a long run. The 1G SerDes still performs clock recovery even when the pluggable module has no CDR.19

The packet path is audio packing, o2at header, Ethernet MAC, 1000BASE-X PCS/PMA and optical conversion. Standard destination and source MAC addresses remain present. There are no IP, UDP, RTP or TCP headers. The MAC supplies the normal Ethernet FCS, preamble, start delimiter and interpacket gap. The PCS supplies 8b/10b coding at 1G or 64b/66b at 10G. These are distinct layers, and the o2at encoder shall not add another private line-code layer.

Laboratory v0.2 frames use EtherType 0x88B5. This is a local experimental EtherType, not a permanent O2AT allocation. Its use shall remain inside the developing organisation’s controlled evaluation domain; a production release needs an appropriately assigned protocol identifier. The lowercase four-byte magic in every packet improves experimental demultiplexing but does not create ownership of the EtherType.20

The minimum interoperable Ethernet MTU is 1,500 bytes. Jumbo frames are not required. Audio payload is capped at 1,440 bytes in v0.2, leaving 48 bytes of unused MTU after the fixed header. Control payload may occupy up to 1,488 bytes. Packets shorter than the Ethernet minimum are padded by the MAC; padding is excluded from the o2at audio-group count and control-body length. Receivers shall validate the complete implied length before accepting data.

A single optional 802.1Q tag may use PCP 6 when the network administrator has mapped that priority appropriately. PCP is a scheduling request, not a reservation or latency guarantee. The default direct link is untagged. Do not send user audio to IEEE bridge-control multicast addresses such as 01:80:C2:00:00:2F. Discover with bounded broadcast traffic on an isolated evaluation link, then use unicast MAC addresses and explicit session admission.

A switch shall be qualified for forwarding the experimental EtherType, the selected VLAN, frame sizes, multicast or broadcast policy and required egress capacity. Network ACLs, port security, unknown-protocol filtering and energy-saving policies can affect operation. No EtherType value guarantees universal NIC driver exposure or switch forwarding. Raw-socket software must likewise be tested against the actual NIC and operating system.

5.4 Fixed 12-byte common header

Offsets below are relative to the beginning of the Ethernet payload, immediately after EtherType. All multi-byte integers are big-endian. Wire version 0x2 occupies the upper nibble of byte 4; it is not a standalone 0x02 byte. The format and channel configuration are immutable session context, established before audio. This header supersedes the inconsistent 18-byte layout in the attached v0.1 draft.

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 6. Every byte in the fixed o2at header
OffsetSizeFieldv0.2 encoding and validation
01 bytemagic[0]0x6f, ASCII lowercase o
11 bytemagic[1]0x32, ASCII 2
21 bytemagic[2]0x61, ASCII lowercase a
31 bytemagic[3]0x74, ASCII lowercase t
41 byteversion_typeBits 7:4 = 0x2. Bits 3:0: 0 = AUDIO; 1 = CONTROL; 2 = FINAL_AUDIO. Other types are reserved and discarded.
51 bytecountAUDIO: group count, 1–180 with 4 × channels × count ≤ 1,440. CONTROL: body length in eight-byte units, valid range 2–186.
6–94 bytessession_idUnsigned receiver-assigned nonzero identifier. Zero is used for discovery and initial OPEN. Binding includes source/destination MAC and VLAN context.
10–112 bytessequenceUnsigned audio packet sequence modulo 65,536, starting at zero for a new session. CONTROL uses zero; transaction IDs are in its body.

Example: a first regular stereo packet with 180 groups and session ID 0x12345678 begins 6f 32 61 74 20 b4 12 34 56 78 00 00. Its 1,440 audio bytes follow immediately. No additional private checksum or payload prefix is inserted. Ethernet FCS covers the frame.

5.5 Control body, capabilities and session configuration

The format/rate/channel map cannot change merely because a new packet arrives. An unknown session, unsupported version, invalid group count, inconsistent length or wrong source binding is rejected before audio FIFO commitment. Ethernet padding is not audio. Frames not matching the magic, supported version and admitted session binding shall be discarded, including unrelated uses of experimental EtherType 0x88B5.

Every CONTROL body starts with a 16-byte prefix. Byte 0 is opcode; byte 1 is kind, with 0 request, 1 response and 2 notification. Bytes 2–3 are a status code, zero for success, requests and notifications. Bytes 4–7 hold a nonzero transaction_id for requests/responses and zero for notifications. Bytes 8–15 hold a source-selected 64-bit session_nonce, zero only during discovery. Prefix fields use big-endian encoding.

TLVs follow the prefix. Each has a uint16 type, uint16 value length, the value and zero padding so that the entire record is a multiple of eight bytes. Length excludes the four-byte TLV header and its padding. Type lookup uses type & 0x7fff; bit 15 marks a critical extension. Reject unknown critical TLVs, nonzero reserved fields, conflicting duplicate mandatory TLVs and any length extending beyond the declared body. Ignore well-formed unknown non-critical TLVs.

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 7. Proposed control descriptors and exact rates
TLV and value lengthExact value encodingPurpose
0x8001 format_descriptor
8 bytes
Byte 0: format 1 = PCM32, 2 = DSD32. Byte 1: channels 2–32. Byte 2: valid_bits, PCM 16/24/32 or DSD 1. Byte 3: zero. Bytes 4–7: uint32 rate_hz.PCM rate_hz is samples/s/channel. DSD rate_hz is one-bit samples/s/channel. No eight-bit multiplier overflow.
0x8002 channel_map
2 × channels bytes
An ordered array of uint16 channel identifiers, in payload order.Stereo is FL then FR. A receiver explicitly accepts the complete ordered map.
0x8003 packet_profile
16 bytes
Bytes 0–1: nominal groups G. 2–3: max outstanding packets. 4–5: prefill packets. 6–7: zero. 8–11: max frame residence in µs. 12–15: max credit interval in µs.G is fixed for regular packets. 1 ≤ prefill ≤ window < 32,768. G and both timing limits must be positive; all grants must fit storage.
0x8004 clock_mode
4 bytes
Byte 0: 0 receiver-paced, 1 established common-reference mode. Bytes 1–3: zero.The mode is a declared system relationship, not an inference from two lock LEDs.
0x8005 credit_state
16 bytes
Bytes 0–7: uint64 grant_end_group. Bytes 8–15: uint64 consumed_groups.Exclusive cumulative send limit and cumulative played/retired group count.
0x8006 total_groups
8 bytes
Unsigned uint64 exclusive final group index.Required in END; may also be supplied at OPEN for a known-length stream.
0x0101 capabilities
12 bytes per record
Each record: format byte, valid_bits byte, two zero bytes, uint32 rate_hz and uint32 channel-count bitmap. Bit n−1 represents n channels.Advertise supported combinations. The mandatory basic stereo PCM modes must be listed.
0x0102 rate_hint
8 bytes
Unsigned Q32.32 groups/s; one PCM sample group or one 32-bit DSD block group.Optional source pacing hint. Credits and actual consumption remain authoritative.
0x0103 clock_status
16 bytes
Byte 0: clock state. Byte 1: flags. Bytes 2–3: zero. Bytes 4–7: nominal external reference Hz, or zero. Bytes 8–11: holdover remaining µs. Bytes 12–15: zero.Mandatory reporting for the reference tier; state/flag meanings and reporting events are defined below.

clock_status states are 0 internal, 1 external acquisition, 2 external locked, 3 holdover and 4 fault. Flag bit 0 means reference signal present, bit 1 means a valid conversion clock, and bit 2 means an external-reference input is supported; other bits are zero. Reference frequency is nominal, not a claim of measured accuracy: use 10,000,000 when the external input is selected, including acquisition and holdover, and zero for internal operation. Holdover remaining is zero outside holdover and is a finite declared limit within it. Reference-tier endpoints include this TLV in successful HELLO, OPEN, START and PING responses and issue a STATUS notification on every state change during an admitted session. HELLO reports the current clock state when no session is active. A valid input signal or lock flag alone does not prove a common source reference.

The initial channel registry assigns 0x0001 FL, 0x0002 FR, 0x0003 FC, 0x0004 LFE, 0x0005 back-left, 0x0006 back-right, 0x0007 side-left and 0x0008 side-right. IDs 0x8000–0x801f identify discrete channels 0–31. Other positions require a published registry entry. A channel-count capability is not permission to reorder a map silently.

PCM rate_hz shall identify an advertised supported value, including 44,100 × 2ⁿ or 48,000 × 2ⁿ for n = 0…6 in the principal v0.2 families. The maxima are 2,822,400 and 3,072,000. DSD uses exact rates 44,100 × M or the explicitly declared 48,000 × M extension, where M is 64, 128, 256, 512, 1,024 or 2,048. DSD2048 is therefore 90,316,800 or 98,304,000, both comfortably within uint32. The exact number is authoritative; a marketing multiplier is not a substitute.

5.6 State machine and receiver credits

The control opcodes are 0x01 HELLO, 0x02 OPEN, 0x03 START, 0x04 END, 0x05 ABORT, 0x06 CREDIT, 0x07 PING and 0x08 STATUS. The receiver moves through IDLE, OPEN, ARMED, RUNNING and DRAINING, returning to IDLE after a controlled end or abort. Link loss or reset retires the session and enters a muted state. Discovery broadcasts should be limited to one per second after an initial bounded attempt; all established-session traffic is unicast.

HELLO uses session ID and nonce zero and returns capabilities. OPEN uses session ID zero with a fresh source nonce and all mandatory requested descriptors. A successful response assigns a nonzero session ID and echoes the complete accepted configuration and nonce. START acknowledges that context and returns initial credit_state. Audio shall not be sent before that acknowledgement. Initial credit shall be at least prefill_packets × G, within validated receiver storage. G shall be at least one and 4 × channels × G shall not exceed 1,440. The credit interval, path delay and source response time must fit the buffer window. Playback begins after the agreed prefill; a stream shorter than that threshold may start once its complete final range is buffered and FINAL_AUDIO plus END agree. A repeated START never resets a running FIFO or sequence counter.

Suggested response codes are 0 OK, 1 UNSUPPORTED_VERSION, 2 UNSUPPORTED_FORMAT, 3 UNSUPPORTED_RATE, 4 UNSUPPORTED_MAP, 5 BUSY, 6 BAD_STATE, 7 BAD_LENGTH, 8 UNKNOWN_SESSION, 9 BAD_TRANSACTION and 10 CAPACITY_EXCEEDED. Unknown major wire versions are discarded at framing; an understood control exchange may additionally report an unsupported requested version or capability.

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 8. Minimum control exchange content
Control exchangeMandatory content
HELLO request / responseRequest: no TLVs required. Response: capabilities (0x0101).
OPEN request / responseRequest: format_descriptor, channel_map, packet_profile and clock_mode. Response: complete accepted context, assigned session ID and echoed nonce.
START request / responseRequest: no mandatory TLVs; header session ID and control nonce identify the admitted context. Response: initial credit_state. No audio before acknowledgement.
CREDIT / STATUS notificationCREDIT requires credit_state; repeat at the negotiated maximum interval. STATUS requires clock_status and is issued when clock state changes.
END request / responseRequest: total_groups. Response: accepted final count, echoed in total_groups. Drain and retirement follow the agreed range.
ABORT or PING request / responseNo mandatory request TLVs. Match opcode, transaction ID and nonce; duplicate requests retain their original effect. A reference-tier PING response also requires clock_status.

Reference-tier HELLO, OPEN and START responses additionally include clock_status. A source may use PING to observe acquisition before issuing START. Failed control responses report their status without implying that an audio context was accepted.

Every response shall use kind 1 and echo the request’s opcode, transaction ID and nonce. CREDIT and STATUS notifications use kind 2 and transaction ID zero. PING shall echo its request transaction. During a running session, valid data or control refreshes liveness; one second without valid peer traffic shall abort and mute. The audio underrun deadline normally acts earlier. A new request uses a new nonzero transaction ID. Retries repeat the identical request and ID; the receiver caches its response so duplicate OPEN, START, END and ABORT requests cannot repeat side effects. Reusing an ID with different contents is an error. A prototype may use a 20 ms initial control retry, exponential backoff and a one-second overall deadline. These startup/control timings are separate from the much shorter negotiated credit-notification interval.

The canonical session key is (talker_MAC, listener_MAC, VLAN context, session_id). Reverse control frames reverse Ethernet source and destination addresses but retain that canonical pair. Audio validates this binding; controls also validate the nonce. OPEN or ARMED sessions expire after two seconds without valid peer activity. PING requests receive matching responses during an otherwise idle established session. Session IDs shall avoid active and recently retired values and shall not restart deterministically after reboot. Random selection or retained allocation state must be combined with collision checking. Retain retired identifiers beyond the admitted frame-residence interval. The 64-bit nonce and 32-bit session field reduce stale-session confusion; they are not authentication or cryptographic replay protection.

The source maintains sent_end_group and may transmit only when sent_end_group + packet_groups ≤ grant_end_group. Grants are cumulative and cannot be retracted. The receiver grants only space that its validated buffer accounting can honour, including audio already in flight. Duplicate or reordered CREDIT notifications cannot issue additional credit; the source ignores older values. The receiver repeats its latest grant at the agreed interval so one lost notification does not cause a permanent stall.

consumed_groups starts at zero and advances with actual group retirement at the playback side. The basic profile aborts and mutes when an expected packet remains missing at its playout deadline, or when invalid data would otherwise enter that active stream. A bounded reorder buffer may accept an early later packet until that deadline; late and duplicate packets are discarded. The source timeline is never compressed to hide a missing packet. OPEN may prepare or acquire the selected clock. START shall succeed only when the conversion clock is valid; common-reference mode additionally requires the established frequency relationship. Loss of clock validity after START requires immediate safe mute and session abort. A holdover transition may retain the session only within its declared clock and FIFO margins; otherwise restart with a new session. When transport remains operational, send a best-effort STATUS with the current session ID and nonce before retirement. Notification delivery shall not delay local protection. STATUS never authorises automatic resumption or a clock-mode change. Optional future concealment would need an explicit negotiated policy and consistent group accounting. Credits pace delivery; they do not change the immutable nominal sample format or promise that a fixed-rate live producer can pause indefinitely.

Every regular AUDIO packet contains exactly negotiated G groups. FINAL_AUDIO contains 1…G groups and is the final packet, including when the final count equals G. END supplies total_groups and acknowledges acceptance of that final count. It may arrive before or after FINAL_AUDIO; acceptance shall not discard preceding in-flight audio. The receiver verifies agreement, drains only that range and closes the session. An empty stream closes with total_groups = 0 and no audio packet. Cached END and ABORT responses remain available after retirement for at least the maximum retry duration plus admitted frame residence. After playing the last group of a validated FINAL_AUDIO packet, the receiver enters a drained, safely muted state and awaits END if it has not yet arrived. No further audio is expected, so the empty FIFO is not an underrun. A missing FINAL_AUDIO packet still invokes the loss policy. ABORT mutes, flushes and retires the session. Restarts require a new nonce/session ID and sequence zero, never a reset within the old session.

The receiver unwraps sequence16 into an extended packet index. For regular packets, the first group index equals extended_packet_index × G. Frames failing protocol or session validation are discarded without disturbing unrelated sessions. A corrupt Ethernet frame may cause a subsequent confirmed gap in the active stream. The active window must remain below 32,768 packets, and admitted frame residence must be shorter than 32,768 / packet_rate. The deployment shall be loop-free and bounded; a 16-bit sequence cannot distinguish arbitrarily old traffic circulating in a network loop.

5.7 PCM and native DSD packing

PCM uses signed two’s-complement 32-bit containers in network byte order. Valid 16- or 24-bit sample values occupy the most significant bits, with unused least significant bits zero. A group contains one container per channel in the accepted map order; groups follow increasing sample time. For 16-bit PCM, +1 is 00 01 00 00 and −1 is ff ff 00 00. For 24-bit PCM, +1 is 00 00 01 00 and −1 is ff ff ff 00. Full 32-bit values have no padding.

DSD32 is a packing format, not 32-bit PCM. Each channel word contains 32 successive one-bit samples, with the earliest sample in bit 31. Emit that word big-endian. The next word belongs to the next channel at the same 32-bit time block, followed by the next group. For a word 0x80000000, only the first one-bit sample is 1; for 0x40000001, the second and thirty-second are 1. No DoP markers, arithmetic scaling or implicit bit reversal are allowed.

“MSB-first” here describes the numerical octet order and the DSD chronological mapping after decoding. It does not override IEEE Ethernet’s bit transmission, scrambling or PCS rules. The MAC/PHY remains standards-compliant. Explicit decoded payload vectors are therefore more useful than describing the laser waveform as an audio word.

5.8 Packetisation and exact occupied-wire budget

Let C be channels and G be groups per packet. Both PCM32 and DSD32 use P = 4CG audio bytes. For a maximum-size audio packet, G = floor(1,440 / (4C)); P is then 1,440 bytes for 2 or 8 channels, and 1,408 bytes for 16 or 32. A 125 µs packetisation deadline additionally limits G to floor(group_rate × 125 µs), with a minimum of one. Use the smaller constraint and negotiate it once per session.

At lower rates, that deadline produces short regular packets rather than variable-duration regular packets. For example, stereo 48 kHz uses six groups and 48 audio bytes per packet. Every bandwidth claim shall specify G or the deadline that determines it. In the extreme-rate table, the maximum-size cap is reached before 125 µs, so all rows use the payload-alignment values above.

For ordinary audio packets here, occupied byte times are P + 50 untagged or P + 54 with one VLAN tag: 12 o2at header, 14 MAC header, optional 4 VLAN, 4 FCS, 8 preamble/SFD and 12 IFG. More generally, occupied byte times are 38 + max(46, 12 + P) untagged, or 42 + max(42, 12 + P) for a minimally padded single-tag frame. Ethernet padding therefore cannot be forgotten for very short packets.

PCM packet_rate = Fs / G. DSD packet_rate = FDSD / (32G). Occupied wire rate = 8 × occupied_byte_times × packet_rate. PCM audio rate = 32 × C × Fs; DSD audio rate = C × FDSD. Compare the occupied rate against 1.000 or 10.000 Gb/s MAC capacity. Do not deduct 8b/10b twice: it is already represented by the 1.25 GBd physical rate supplying the 1G MAC service.

The advanced direct-link ceiling is 85% of occupied wire capacity, including other traffic. Eight-channel 48-family extremes use 81.59232%, leaving 18.40768% physical capacity and 3.40768 percentage points below that policy. An 80% policy rejects this mode. The bandwidth budget alone does not establish a switched-latency bound.

5.9 High-rate capacity and admission

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 9. Extreme-rate budgets, including all occupied-wire overhead
Equivalent payload formatsCh.G / audio bytesAudio Mb/sWire Mb/s untaggedWire Mb/s taggedInterval µsAdmission
PCM2,822.4
DSD2048/44.1
2180 / 1,440180.6336186.9056187.407463.77551G admitted
PCM3,072.0
DSD2048/48
2180 / 1,440196.6080203.4347203.980858.59381G admitted
PCM2,822.4
DSD2048/44.1
845 / 1,440722.5344747.6224749.629415.94391G admitted
PCM3,072.0
DSD2048/48
845 / 1,440786.4320813.7387815.923214.64841G admitted
PCM2,822.4
DSD2048/44.1
1622 / 1,4081,445.06881,496.38521,500.49057.794810G required
PCM3,072.0
DSD2048/48
1622 / 1,4081,572.86401,628.71851,633.18697.161510G required
PCM2,822.4
DSD2048/44.1
3211 / 1,4082,890.13762,992.77033,000.98093.897410G required
PCM3,072.0
DSD2048/48
3211 / 1,4083,145.72803,257.43713,266.37383.580710G required

PCM entries are kHz; DSD entries identify the underlying family. PCM32/2,822.4 kHz and DSD2048/44.1 have identical payload rate. PCM32/3,072 kHz and DSD2048/48 also match. The 48 kHz-derived DSD form is explicitly an extension. Values are decimal Mb/s; intervals are packet periods, not complete audio latency.

5.10 Earlier high-resolution formats and useful scaling

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 10. Lower-rate examples with the same 125 µs packetisation policy
FormatChannelsG / audio bytesAudio Mb/sTagged wire Mb/s1G utilisation
PCM32 / 768 kHz296 / 76849.152052.60805.261%
PCM32 / 768 kHz845 / 1,440196.6080203.980820.398%
PCM32 / 1,536 kHz2180 / 1,44098.3040101.990410.199%
PCM32 / 1,536 kHz845 / 1,440393.2160407.961640.796%
DSD512 / 44.1288 / 70445.158448.62234.862%
DSD512 / 44.1845 / 1,440180.6336187.407418.741%
DSD1024 / 44.12176 / 1,40890.316893.78079.378%
DSD1024 / 44.1845 / 1,440361.2672374.814737.481%
DSD1024 / 482180 / 1,44098.3040101.990410.199%
DSD1024 / 48845 / 1,440393.2160407.961640.796%

Stereo PCM32/768 kHz has 49.152 Mb/s audio payload but exceeds 50 Mb/s once the stated framing and packetisation are included. This corrects the loose claim that every such stream consumes less than 50 Mb/s on the wire. At the upper extreme, stereo still uses only about one fifth of 1G, so the baseline decision remains strong.

A 1G endpoint can also carry many channels at lower rates. For example, 32 channels of PCM32/768 kHz use approximately 816.5935 Mb/s with the aligned 1,408-byte payload and VLAN tag. Sixteen or 32 channels at PCM3072 or DSD2048 exceed 1G and justify the optional 10G profile, even when all traffic remains Layer 2. Mixed-speed fabrics shall admit the sum of streams at every egress port.

6. Clocking Architecture, 10 MHz Discipline & Flow Control

6.1 Separate serial recovery from conversion timing

An SFP receiver and host SerDes recover enough serial timing to identify symbols and bytes. That recovered clock is necessary for transport, but it shall not directly clock the D/A conversion process in any O2AT receiver. The reference tier additionally requires external-reference hardware. Accepted packets enter a dual-clock FIFO. The write side belongs to the MAC or local transport domain; the read side belongs to the audio domain. Packet-delay variation changes FIFO occupancy while the conversion cadence remains controlled locally. The Ethernet SerDes and MAC clocks shall independently satisfy their Ethernet timing requirements; disciplining the audio clock does not imply that Ethernet or USB signalling clocks are controlled by the 10 MHz input.

This architecture removes a direct timing transfer path. It does not make all interface effects disappear. Supply noise from the optical receiver can still modulate an oscillator, inappropriate clock-domain crossings can corrupt data, and inadequate buffering can cause discontinuities. A defensible claim is that packet arrival is not the conversion-clock reference, subject to a validated buffer and clock-control design. An unqualified claim of zero interface jitter or automatically superior sound is not warranted.

Figure 1. Receiver clock authority and the return control path

Audio packets pass from source through optical Ethernet and a validated packet buffer to a dual-clock FIFO. A locally generated audio clock drives FIFO reads and conversion. Receiver credits return to the source over the second fibre.

The optical receive clock supports byte recovery and FIFO writes. The audio clock supports FIFO reads and the converter interface. Feedback controls delivery from the source; it does not substitute packet arrival for the conversion reference.

6.2 Why the reference tier includes a 10 MHz input

A shared reference can make the source and sink sample-rate generators frequency coherent. This is useful when a source has a real-time acquisition or DSP clock that cannot be freely paused, when several receivers must avoid long-term differential drift, or when a system already distributes a high-quality reference. It can remove the need for a network-time servo in a dedicated installation. For that reason, an external 10 MHz input is mandatory hardware on an o2at-reference receiver and optional on o2at-basic.

The benefit is conditional on the actual audio clocks being derived from that reference with the intended rational ratios. Merely connecting two BNC sockets to the same oscillator does not prove that their audio paths are disciplined. A common reference suppresses relative frequency error in a properly locked design; it does not by itself align divider phase, sample number, stream start or DAC group delay. Different startup phases can leave a fixed offset even when long-term rates agree.

The earlier phrase “zero-ppm-drift synchronous FIFO operation” is therefore an idealised frequency result, not a complete synchronisation method. Real lock acquisition, reference faults, holdover drift and servo behaviour require measurement and status reporting. Multiple DACs needing sample-accurate acoustic alignment require an explicit epoch or trigger mechanism and calibration beyond v0.2’s consumer link profile. PTP or another time-distribution mechanism remains appropriate where network-wide time is actually required.

Frequency accuracy and short-term phase noise are also different specifications. A highly accurate 10 MHz source can produce a worse audio clock after an unsuitable synthesiser than a good local oscillator. The reference design should compare phase-noise spectra and integrated jitter over stated offset bands in internal and external modes, with repeatable conditions. A narrow reference-tracking loop and a suitable local oscillator may preserve short-term noise performance while following the long-term reference.

6.3 Clock-input and synthesiser engineering

The proposed reference input is a labelled 50 Ω BNC port. Its baseline qualification shall include a 10 MHz sine input at 0.5 V rms into 50 Ω, approximately +7 dBm, and an AC-coupled 50% square input of 2 V peak-to-peak at the termination. The square-wave qualification is zero-mean at the terminated input, with no DC offset applied. The manufacturer shall state accepted amplitude ranges, input coupling, absolute limits and lock thresholds. A 50 Ω port must not silently be specified as 75 Ω, and voltage, power and waveform conventions must be explicit.

The front end should use appropriate termination, filtering, protection and a low-additive-noise conditioning stage. It shall detect missing or out-of-range reference input. Entering or leaving external lock shall trigger a controlled transition: report status, stop or drain the stream where necessary, mute safely, reacquire and restart. Switching an audio clock source while the converter is live is not an acceptable substitute for a transition design.

Si5341 and LMK04828 are examples requiring different clock plans, not interchangeable generic fractional-N audio solutions. Si5341 is a flexible clock generator using fractional MultiSynth synthesis; its presence alone does not establish that it attenuates input jitter as required. LMK04828 is a dual-loop jitter-cleaner/distribution device whose oscillator, PLL and divider constraints must be checked against the desired audio rates. The usual LMK04828 dual-loop cleaning plan also requires an external VCXO or tunable crystal. Neither part should be selected from its headline jitter number without analysing the full reference-to-output transfer.21

The 44.1 and 48 kHz families shall have explicit frequency plans, including reference division, oscillator choice, output division, spur locations, lock time and output interface. At the new extreme rates, familiar 256 × Fs assumptions are often inappropriate. Use the converter’s supported clock architecture rather than multiplying a legacy ratio until it exceeds the device’s limits. FPGA PLLs can serve PCS and internal logic, while the conversion clock should be provided by a measured clock subsystem appropriate to the DAC.

A shared coaxial clock distributor may reconnect chassis electrically through BNC shields even when the audio path is optical. Therefore, O2AT guarantees a non-conductive audio-link path, not unconditional whole-system galvanic isolation. A reference installation must account for those parallel paths and any isolation implemented by the clock inputs or distribution equipment.

6.4 Three rate-control cases

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 11. Rate control is an end-to-end property
CaseWho controls sample production?Required behaviour
Stored or generated playback with elastic sourceReceiver clock sets consumption; source can fetch or generate on demandUse receiver credits and bounded source buffering. The source shall stop advancing into the link when grant is exhausted.
Actual source and DAC clocks share a coherent referenceBoth sample clocks follow the same valid frequency planObey cumulative credits at all times; verify long-term occupancy. Shared frequency does not define sample epoch or authorise a burst beyond its grant.
Independent fixed-rate live sourceSource clock cannot pause and is not locked to the DACA finite FIFO eventually fails under sustained mismatch. Lock a real clock, use explicit rate conversion, or reject the transparent mode.

A legacy S/P-DIF or ADAT input is an important fixed-rate example: its stream continues according to the sending device’s clock and has no o2at credit interface. An active bridge must establish a common clock, control the true source where possible, or use an explicitly selected rate-conversion path. A future source-clock-tracking profile would need its own estimator, servo and jitter-transfer specification. Neither buffering nor PAUSE alone makes an arbitrary unlocked legacy source transparently synchronous with a local DAC.

IEEE 802.3x PAUSE is a supplementary direct-link safety mechanism. It asks the adjacent transmitting MAC to suspend data transmission, using a 16-bit pause-time value in units of 512 bit times. It is not an audio-rate feedback message and is not forwarded as an end-to-end command by ordinary bridges. On a switched path it controls the neighbouring switch port, not the original audio producer. It can also block unrelated traffic on that link.22

PAUSE may help an elastic, unlocked source avoid a transient receive overflow, provided it responds in time and has upstream storage. It cannot repair indefinite drift from a free-running live source, make an isochronous USB stream retransmittable or prevent a receiving FIFO from draining while a source is paused. The normative o2at receiver-credit mechanism therefore remains necessary; a USB bridge must propagate demand to the USB audio feedback mechanism as well.

6.5 Buffer sizing, latency and loss behaviour

Buffer sizes shall be expressed in bytes and audio groups with an associated format, not simply in “samples”. At stereo PCM32/3,072 kHz, 16,384 sample groups occupy 131,072 bytes and represent 5.333 ms. Eight channels require 524,288 bytes for the same group count. For DSD, a group spans 32 one-bit times per channel, so the group rate is the DSD bit rate divided by 32.

For an uncompensated relative rate error ε, the approximate time to consume an available FIFO margin of M groups is M / (group_rate × |ε|). As an illustration, a 1 ms margin with 50 ppm mismatch lasts about 20 s. A large FIFO postpones a steady-state problem; it does not eliminate it. Margin must also cover in-flight packets, source response time, processing stalls and any admitted switch bursts.

An initial reference design may target 0.5–2 ms of FIFO occupancy, but this is a prototype operating range, not a universal latency promise. Define the packetisation deadline, the largest allowed burst, the credit update interval and the worst-case control round trip, then size free space above the high watermark accordingly. Credits shall count data already sent but not yet consumed so that outstanding packets cannot overrun the receiver.

The latency budget includes packet assembly, serialisation, fibre propagation, switch forwarding/queueing if present, FIFO residence and the converter interface. A full 1,440-byte audio packet occupies 11.952 µs of tagged 1G wire time including IFG. That is only one component of the path. At low rates, a 125 µs packetisation deadline avoids waiting 3.75 ms to fill a stereo 48 kHz packet. Shorter packets trade bandwidth efficiency for assembly latency.

Ethernet FCS detects frame corruption; it is not forward error correction and there is no mandatory retransmission of late audio. Frames failing FCS or syntax validation are discarded and counted. A confirmed gap, invalid active-stream state or underrun shall invoke that stream’s defined mute/restart procedure; unrelated malformed or foreign traffic shall not mute an admitted stream. PCM may be ramped towards zero by an explicitly identified output-protection stage. Raw DSD shall not be replaced blindly by an all-zero bitstream, which represents a large DC condition in the one-bit domain. Use converter mute or a validated DSD idle/transition sequence.

6.6 Exact clock ratios and acceptance measurements

For a nominal 10,000,000 Hz reference, the following ratios are exact arithmetic frequency targets. They are not validated register configurations for any particular synthesiser. In both families, a 32-bit DSD group advances at the corresponding maximum PCM sample-group rate.

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 12. Exact frequency relationships for clock-plan verification
Clock targetFrequency, HzExact ratio to 10 MHz
PCM sample rate, 44.1 family2,822,400882 / 3,125
PCM sample rate, 48 family3,072,000192 / 625
DSD2048 bit clock, 44.1 family90,316,80028,224 / 3,125
DSD2048 bit clock, 48 extension98,304,0006,144 / 625

A published clock plan shall identify reference frequency, programmed ratios, oscillator and divider ranges, output electrical interface, acquisition time and fault policy. Exact rational arithmetic removes nominal ratio error. It does not remove reference frequency error, residual loop noise, fractional spurs or startup phase ambiguity. Common frequency discipline is therefore assessed separately from a shared sample epoch.

Clock acceptance shall report output frequency and uncertainty, relative drift between source and sink, phase-noise spectra with stated integration limits, discrete spurs, and time-domain period or phase error. Test both rate families, internal and external reference modes, optical idle and maximum admitted traffic, temperature variation, reference interruption and reacquisition. Absolute frequency accuracy requires an independent measurement reference; comparing two clocks disciplined by the same source only establishes their relative behaviour.

Log FIFO occupancy and cumulative group counts during the same tests. Common-reference operation shall demonstrate bounded occupancy without sample insertion or deletion. Receiver-paced operation shall demonstrate stable upstream rate control across the admitted delay and oscillator-error envelope. Declare holdover duration and drift bounds; a fixed-rate transparent stream shall stop safely before its remaining FIFO margin can be exhausted. A synthesiser lock indication does not replace these measurements.

7. Technical Challenges & Reference Implementation

7.1 Power, RF coupling and clock-domain crossings

Place the optical cage and high-speed PHY circuitry away from the lowest-noise oscillator, DAC reference and analogue output paths. Use a controlled return-current structure, appropriate cage bonding, local regulation and filtering, adequate thermal paths and a layout validated for the actual serial channel. Galvanic isolation between boxes does not isolate a DAC’s local analogue circuitry from its own SFP supply current.

Select optics by supported rate, optical budget, qualification, diagnostics and measured electrical behaviour. A non-retimed low-power module can be a sensible reference choice, but CDR absence is not a universal conformance requirement. Faster optics do not necessarily corrupt conversion timing, and a nominally slower optic is not automatically quieter. Test supply spectra and analogue output under realistic payload, burst, idle, module temperature and cable-loss conditions.

A dual-clock FIFO should use independently clocked read and write pointers, Gray-coded pointer transfer, synchronisers with a documented metastability analysis, and correctly constrained reset release. Do not synchronise a changing binary multi-bit counter by placing two flip-flops on each bit. Data storage, metadata and format changes must cross as a coherent transaction. Vendor FIFO IP can help, but its configuration and timing constraints still require review.

The receive path shall not release an audio packet to the playback FIFO until the complete frame has passed FCS and syntax checks. Either buffer the whole frame before committing it or support an atomic rollback that never exposes corrupt groups. A late CRC result cannot undo samples already converted. Likewise, a format change shall drain or discard the old stream under mute before a new session is admitted.

7.2 FPGA and host reference design

A practical first platform is Sipeed’s Tang Mega 138K Pro, which provides two SFP+ cages and a GW5AST-LV138FPG676A device with transceivers spanning rates that include 1.25 Gb/s. The baseline experiment uses a qualified 1G optic in a compatible cage and configures the FPGA accordingly. It does not require operating the optical link at 10G merely because the board has SFP+ cages.23

The project must implement or integrate the 1000BASE-X PCS/PMA, MAC, receive validation, control-state machine, audio packing, dual-clock FIFOs and output interface. Verify the available transceiver reference clocks and rate configurations before board commitment. The board’s commercial Gowin toolchain requirement and silicon-revision restrictions must be recorded. Open RTL does not imply a fully open implementation toolchain or a ready-made, validated O2AT bitstream.

A host prototype can send raw Ethernet frames through a suitable NIC, but an AF_PACKET socket does not automatically provide zero-copy, bounded scheduling latency or reliable PHY control. Measure the actual path, use appropriate batching or ring buffers where supported, and keep audio buffering and receiver credits visible. The release should publish a packet dissector, a reference encoder/decoder, deterministic test vectors and reproducible builds before attempting broad hardware optimisation.

The optional expansion profile uses 10GBASE-R at 10.3125 GBd and a 10.000 Gb/s MAC service. Qualified 10GBASE-SR or LR media are allowed under an explicit profile. A mixed-speed switch can bridge a 10G source to a 1G DAC port only after admission to the slower egress. Mixed-speed paths shall also bound source or upstream-switch bursts to the qualified egress service and queue capacity. Average admission and receiver credits alone do not protect an intermediate switch from a rapid burst at 10G. The protocol remains raw Layer 2; no IP workload is required to justify a 10G link.

7.3 USB ingress and DAC egress impose separate ceilings

The conventional high-speed USB 2.0 high-bandwidth isochronous endpoint limit is three transactions of up to 1,024 bytes per 125 µs microframe. The maximum payload is therefore 24.576 MB/s, or 196.608 Mb/s. Stereo PCM32/3,072 kHz requires exactly that nominal payload, leaving no positive rate margin for asynchronous clock error. A 385-sample-group microframe would need 3,080 bytes. The 44.1 kHz-family extreme fits the theoretical three-transaction budget, but endpoint and driver support remain separate requirements.24

A USB-to-o2at bridge also needs a real feedback controller. Ethernet credits prevent receive-buffer overrun, but do not by themselves specify UAC2 feedback. Estimate downstream consumption from cumulative DAC-side group counters against the bridge’s observed USB microframe count, filter bounded report-delay variation and control source-buffer occupancy. Define the conversion from o2at groups to the advertised USB sample units; PCM and vendor-specific native-DSD endpoints need different mappings. Verify loop stability, startup, positive/negative clock offsets and feedback quantisation. Packet interarrival variation shall not become the DAC conversion-clock reference.

Generic “XMOS UAC2” is not enough evidence for these rates. As a concrete documented implementation limit, XMOS lib_xud v4.0.1 supports at most two transactions per microframe, on the specified xcore.ai configuration. That yields 16.384 MB/s and falls short of stereo DSD2048 or PCM32 at 2,822.4/3,072 kHz. This is a limit of the cited implementation, not a statement about every possible future XMOS design.25

Begin extreme-rate transport validation with a native FPGA pattern generator or a separately qualified high-throughput ingress. A suitable USB 3 implementation is a possible later bridge. The intended user-facing format must be supported by the source software, host driver, endpoint descriptors, device firmware and downstream interface. A successful optical packet test is not evidence that a present-day music server can generate the same format.

At the DAC side, two 32-bit I²S slots require BCLK = 64 × Fs. PCM32/2,822.4 kHz therefore needs 180.6336 MHz BCLK, and PCM32/3,072 kHz needs 196.608 MHz. One serial DSD line per channel needs 90.3168 or 98.304 MHz at the extreme rates. These are calculations from the stated packing, not guaranteed operating modes of existing I²S-over-HDMI links. Multiple internal lanes or a converter-specific interface may be necessary.26

7.4 Applying the architecture to the N30SA, µDDC and Harmony

Aurender explicitly states that an external clock input on its source does not discipline asynchronous USB audio; the downstream audio device controls average data flow. Thus, connecting an N30SA by USB to a µDDC does not make the µDDC the N30SA’s USB signalling-clock generator. In asynchronous USB operation, the sink’s feedback controls the host’s average delivery of audio samples. The µDDC’s internal relationship between that feedback and its external reference remains undocumented in the public specifications reviewed here. The N30SA’s reference-clock treatment of AES/S/P-DIF is a different path.27

LAiV specifies a 10 MHz, 50 Ω clock input for the Harmony µDDC, accepting square or sine waves, and lists PCM through 768 kHz and DSD512 on its USB/I²S paths. Those specifications do not establish support for DSD2048 or PCM3072, nor do they disclose every internal feedback and conversion-clock relationship. The original Harmony DAC in this experiment has no 10 MHz input, so reference discipline would be applied at the intermediate interface rather than directly at that DAC.28

The useful first experiment is therefore a transparent bridge within the existing equipment’s supported formats: asynchronous USB ingress, o2at optical transport, a receiver with explicit local or disciplined read clock, and a validated µDDC/DAC-compatible output. Confirm where the downstream conversion clock is actually selected. Separate transport stress tests at DSD2048/PCM3072 can then demonstrate future capacity without implying that the current playback chain accepts those rates.

8. Device Labelling, Diagnostics & User Experience

An O2AT socket should be as clear to use as the TOSLINK socket it replaces. The label should identify an audio interface and its optical class. It should not imply that any similarly shaped SFP port on a switch, USB extender or proprietary DAC is interchangeable. A direct connection should expose only useful outcomes: link established, compatible format, reference state and any fault requiring attention.

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 13. Product-facing information with engineering meaning
Marking or diagnosticRequired interpretation
o2at 1G / SX / LC duplexStandard baseline optical profile; matched 850 nm multimode optics and fibre
o2at-reference / 10 MHz / 50 ΩExternal clock input present, with the electrical acceptance limits in the manual
PCM32 maximum rate; DSD rate and familyReport verified endpoint capabilities, including the 44.1 or 48 kHz DSD family
Link / session / clock stateSeparate physical link-up, negotiated session and conversion-clock lock; one green light must not hide an unsupported format
CountersFCS rejection, sequence gaps, syntax errors, FIFO extrema, underrun/overrun, reference loss and restart count
Measured latency modeReport the configured packetisation deadline and FIFO target, plus the qualified topology, rather than a generic “zero jitter” claim

After discovery, the source shall present only mutually supported modes and refuse an over-budget channel/rate combination. Reference loss should produce a clear status with a documented holdover or controlled stop policy. A user should not need to choose endian order, DSD bit direction or an HDMI pin map. Those choices belong to the standard and the endpoint implementation.

Diagnostic access may be local, through a companion control application or via an optional management interface. It shall not be necessary for basic direct playback. Optical O2AT does not carry electrical power between devices. Any separate powering arrangement is outside the optical audio profile and must not be advertised as inherent optical PoE.

9. Industry Adoption, Validation & Governance

9.1 Why adoption will be gradual

Manufacturers already have substantial investment in USB receivers, XMOS firmware, S/P-DIF input circuits, certification, drivers and support processes. Those are working product assets. A new optical socket adds a cage, an Ethernet-capable digital path, firmware, power and thermal requirements, plus another interoperability matrix. Conservative adoption is a rational response to those costs, especially while compatible sources are scarce.

Consumers recognise TOSLINK and USB and can connect them with inexpensive cables. O2AT must therefore win on clear practical benefits: isolation, a documented format envelope, robust negotiation, replaceable standard optics and dependable operation. Extreme sample rates provide design headroom, but do not by themselves establish a listening benefit or persuade a manufacturer to redesign a successful DAC.

Boutique partnerships can create the first compatible source module, receiver board and reference DAC input. A small common baseline is more useful than several incompatible demonstrations. Open reference RTL, schematics, conformance traces and an accessible analyser give independent engineers a way to reproduce behaviour and identify defects without relying on brand-specific claims.

9.2 A measurable validation programme

  • Protocol: publish exact PCM and DSD test vectors; validate packing, channel maps, sequence wrap, session restart, malformed lengths, unknown critical TLVs and experimental EtherType coexistence.
  • Link and thermal: test qualified optics over the declared loss and temperature range, including module replacement, loss of signal, fault recovery, rail transients and sustained maximum-rate traffic.
  • Rate and buffering: run both frequency families, stereo and eight-channel extremes, bursty hosts, bounded switch contention, controlled ppm offsets and clock-source changes. Log long-term FIFO occupancy and grants.
  • Clock and analogue: measure phase-noise spectra and integrated jitter with stated bandwidths; compare converter output under internal/external reference, link idle/full traffic and alternative qualified modules.
  • Interoperability: demonstrate two independent source implementations and two independent receivers, plus a documented switch configuration for the optional fabric profile.
  • Failure behaviour: inject missing, duplicated, stale, out-of-order and corrupted frames; verify counters, bounded reaction, safe mute and recovery without unintended full-scale PCM or DSD output.

For an illustrative error-rate check, zero observed errors over N independent tested bits gives an approximate one-sided 95% upper bound of 3/N on the bit-error probability. A 24-hour test at 1.25 Gbit/s exercises about 1.08 × 10¹⁴ serial bits, giving an illustrative bound near 2.8 × 10⁻¹⁴ under that model. A MAC-only payload test measures a different population and should be labelled accordingly. Neither result proves zero future errors or audible superiority.

Likewise, a hypothetical BER of 10⁻¹² at 1 Gbit/s corresponds to a mean raw-bit error interval of about 1,000 s, not eleven days. Packet error statistics, detected/rejected frames and actual audio interruption rates are distinct quantities. Publish exposure time, tested bit population, confidence model and failure policy rather than a single reassuring number.

9.3 Release gates and future extensions

Version 0.2 should remain an engineering draft until the packet/control definitions have two independent implementations, the baseline optical profile has qualified parts and the conformance suite exposes no unresolved safety-of-output failures. Public commercial deployment also requires a proper protocol identifier. A governance group should manage the capability registry, change process, errata and reference releases with public decisions.

Future work may add sample-accurate multi-DAC epochs, redundant paths, authenticated control, scheduled Ethernet operation and broader sample formats. These should use negotiated extensions and explicit version changes where necessary. The mandatory consumer baseline should remain small enough to implement without a full IP stack, PTP servo or proprietary licence. The standard’s success should be measured by interchangeable endpoints and reliable playback, not by its longest list of optional features.

9.4 Engineering review status and evidence required

This revision has been checked for format arithmetic, protocol consistency, clock-domain assumptions and alignment with cited primary documentation. It is an engineering proposal suitable for external technical review, not a completed independent peer review or a measured prototype report. The distinction matters most for conversion-clock performance and deterministic latency, which cannot be certified from the packet format alone.

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 14. Peer-review questions and release evidence
Review itemStatus in v0.2Required evidence before a conformance claim
1G SFP capacity through DSD2048 / PCM3072Analytically established for the stated packetisation and channel countsBit-exact sustained tests, measured packet loss and admitted burst behaviour
Wire format and control semanticsExplicit draft definitions with resolved length, session, credit and final-packet rulesIndependent implementations, published vectors and fault-injection results
10 MHz to audio clock planFrequency relationships and measurement requirements defined; no universal component solution assertedDevice-specific synthesis/divider plan, lock/holdover traces, spur and phase-noise measurements
Isolation and endpoint noiseOptical path is non-conductive; other installation paths remain relevantComplete grounding review, rail/noise measurements and DAC-output comparison
Latency and clock feedbackBudget components and control requirements identifiedWorst-case measurements for the admitted topology and a stable USB/receiver feedback loop
Legacy TOSLINK and future ADAT migrationPCM/DSD transport available; broader legacy profiles explicitly scopedMetadata mapping, format negotiation and clock-authority tests on real legacy equipment

Appendix A. Commercial Pushback from the Audiophile Cable Industry

A commodity optical interface challenges the economics of premium metallic digital interconnects. The original proposition compared USB, AES/EBU, S/P-DIF and silver I²S-HDMI cables priced at US$500 to US$5,000 or more with an US$8 duplex-LC patch cord and US$15 SFP modules. In this revision those figures are illustrative commercial assumptions, not verified current offers or a guaranteed system bill of materials. Two endpoints, their optics, electronics, integration and compliance work must all be counted.

A.1 What commodity fibre changes physically

A non-conductive glass-fibre signal path cannot carry a chassis-to-chassis ground current along the fibre. A continuous metallic conductor can. Silver purity, conductor geometry and shielding do not turn that conductor into glass. However, transformer-coupled electrical interfaces and active isolation products can provide galvanic barriers of their own, and ordinary TOSLINK already provides optical isolation. The defensible comparison is the complete interface topology, not a claim that every expensive cable is equivalent or every copper-based system lacks isolation.

A correctly specified inexpensive optical patch cord can therefore deliver the relevant isolation property without precious-metal conductors. Its quality still matters for insertion loss, connector polish, mechanical durability, bend handling and reliable mating. O2AT does not eliminate engineering quality; it relocates much of the relevant performance from luxury conductor construction to standard optical specifications and endpoint design.

A.2 Likely incentives, not an allegation of coordinated resistance

Some firms whose margins depend on premium metallic digital cables may have little incentive to promote an interchangeable optical connection. Others may sell qualified fibre assemblies, rugged connectors, installation services, optical diagnostics or complete interface products. These are plausible commercial responses, not evidence of a coordinated campaign by any manufacturer. No particular firm’s motives can be established from its product price.

The strongest adoption argument is an open and measurable interface. Demonstrate bit integrity, controlled failure behaviour, real isolation boundaries, reproducible clock measurements and an honest complete cost. Listening comparisons can supplement that evidence when appropriately controlled, but they cannot substitute for the transport specification or prove that an expensive digital cable is necessary for a correctly functioning link.

Engineering claims and their limits

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 15. Claims that should be tested or qualified
ClaimEngineering response
A better optical cable must change the decoded sample valuesWithin the operating margin, correctly received values are the transmitted values. Diagnose loss and detected errors rather than attributing a new numerical resolution to the cable.
Optical isolation guarantees no noise anywhere in the DACIt removes the conductive signal path between boxes. The receiver’s own optics, power system and digital logic still generate local activity.
An external clock always improves conversionCompare the complete reference-to-audio-clock chain. Frequency accuracy, phase noise, spurs and clock-selection behaviour are separate measurements.
1G is superior to 10G on every relevant metric1G is the preferred low-power consumer baseline. 10G has real advantages in aggregate capacity, higher channel counts and serialisation time.
A boutique cable business must resist an open optical standardCommercial incentives vary. The standard should succeed through interoperability, test evidence and useful products, without speculation about individual firms.

Appendix B. Glossary of Acronyms and Engineering Terms

This glossary is the terminology companion to the source bibliography. Definitions are scoped to the proposed O2AT interface and distinguish optical hardware, Ethernet transport and audio timing.

This table is wider than the reading column. Scroll horizontally to read all columns.

Table 16. Acronyms and terminology
Acronym or termMeaning in this document
1000BASE-X / SX / LXX identifies the 1G PCS family; SX is the short-wavelength multimode optical profile and LX is a longer-wavelength optical profile. Compatible PHY, module and fibre choices are required.
10GBASE-R / SR / LRR identifies the 64b/66b 10G PCS family; SR and LR identify short- and long-reach optical profiles. They do not define an audio format.
8b/10b; 64b/66bLine codes used here by 1G and 10G Ethernet respectively. Their overhead explains why serial signalling rates exceed MAC service rates.
ADAT / S/MUXADAT Lightpipe is a multichannel optical framing system. Sample multiplexing carries higher sample rates by consuming additional channel slots; a future O2AT bridge needs an explicit clock policy.
AES3 / AES10 / AES67Professional stereo digital audio, MADI multichannel audio, and audio-over-IP interoperability standards respectively. They solve different interface problems.
AoIPAudio over Internet Protocol. Media carried through an IP network, commonly with UDP and RTP in interoperable professional audio systems.
Asynchronous USB audioUSB audio operation in which sink consumption is independent of USB bus timing and feedback controls the host’s average delivered sample rate.
AVB / MilanAudio Video Bridging is an IEEE family of timing and traffic-management mechanisms. Milan defines an interoperable professional media profile built on AVB.
BCLK / LRCLK / MCLKBit clock, left/right or word-select clock, and master clock. Their frequencies and relationships are determined by the converter interface, not simply by a connector.
BERBit-error ratio or probability under defined conditions. A measurement needs a tested bit population, exposure and confidence model.
BiDiBidirectional optical module arrangement using different wavelengths on a fibre. It is distinct from this baseline’s two-fibre duplex-SX arrangement.
Big-endianA multi-byte number is represented with its most significant byte first. This describes packet content, not a redefinition of Ethernet physical bit order.

Glossary of acronyms and terms (continued)

This table is wider than the reading column. Scroll horizontally to read all columns.

o2at table 18
Acronym or termMeaning in this document
BMCBi-Phase Mark Coding. Self-clocking coding associated with conventional IEC 60958/AES3 streams.
BNC / impedanceA coaxial connector family. A stated 50 Ω or 75 Ω impedance is a property of the complete interface and termination, not a cosmetic label.
CDRClock and data recovery. Recovers serial timing for symbol decisions; it is distinct from the clock driving D/A conversion.
CDCClock-domain crossing. Transfer between unrelated or independently phased clock domains, requiring a defined metastability and data-coherency strategy.
Channel groupIn o2at PCM, one 32-bit container per channel at a sample instant. In DSD, one 32-bit chronological block per channel.
Clock disciplineControl of an oscillator or synthesiser against a reference. Frequency discipline alone does not identify a shared sample epoch.
Clock epochA common definition of sample zero or another time origin. Required for aligned start across independent converters; not supplied by frequency lock alone.
CreditA receiver-issued cumulative permission to transmit up to an exclusive group index. A duplicate cumulative grant must not allocate extra space.
DAC / DDCDigital-to-analogue converter; digital-to-digital converter. A DDC can change interface format while leaving the downstream conversion clock architecture product-specific.
Deterministic latencyA latency bound established under explicit traffic, topology and implementation assumptions. It is not implied by Layer 2, an EtherType or a fast link alone.
DoPDSD over PCM framing. Lossless carriage of unchanged DSD bits in PCM-shaped containers with identifying markers.
DSD / DSD2048Direct Stream Digital, a one-bit audio representation. DSD2048 normally means 44,100 × 2,048 = 90.3168 MHz per channel; the 48 kHz-derived extension is explicitly labelled.
DSD32The o2at packing of 32 chronological DSD bits in one big-endian word per channel. It is neither 32-bit PCM nor a DSD rate multiplier.
Duplex LCTwo LC optical connector paths, one in each direction for the baseline link. Duplex describes directionality and fibre arrangement, not Ethernet compatibility.

Glossary of acronyms and terms (continued)

This table is wider than the reading column. Scroll horizontally to read all columns.

o2at table 19
Acronym or termMeaning in this document
EMI / RFElectromagnetic interference and radio-frequency activity. Coupling depends on physical paths, spectra, impedances and the receiving circuitry.
EtherTypeThe Ethernet field identifying the encapsulated protocol. 0x88B5 is for local experimental use, not a permanent O2AT allocation.
FCS / CRCFrame check sequence and cyclic redundancy check. Ethernet uses its FCS to detect corruption; this is not audio forward error correction.
FECForward error correction. Additional redundancy allowing some errors to be repaired. It is not provided merely by an Ethernet FCS.
FIFOFirst-in, first-out storage. A dual-clock FIFO separates write and read domains but requires a long-term rate-control strategy.
FPGA / RTLField-programmable gate array; register-transfer-level design description. Open RTL does not automatically make the implementation tools or third-party IP open.
Galvanic isolationAbsence of a conductive connection across a specified barrier. A fibre signal path can be isolated while other installation connections still join chassis.
GBd; Gb/s; MB/sGigabaud, gigabits per second and megabytes per second. Decimal units are used in the capacity calculations; one byte is eight bits.
Gray codeA code in which adjacent count values differ by one bit. Used with suitable synchronisation and constraints for FIFO pointer crossing.
HoldoverOperation from a local clock model after the external reference is lost. Drift and allowed duration must be specified and monitored.
IEC 61937Encoded non-linear-PCM audio carriage over the IEC 60958 interface family. Its burst and metadata semantics require an explicit legacy-carriage profile.
I²S / LVDSInter-IC Sound serial bus; low-voltage differential signalling. An HDMI-shaped I²S connector does not define a universal external pinout.
IFGInterpacket gap, also called interframe gap. The minimum idle interval consumes link time and is included in the occupied-wire calculation.
ISI / modal dispersionInter-symbol interference; spreading caused by different optical propagation paths. Link impairment depends on fibre, launch, length and receiver behaviour.

Glossary of acronyms and terms (continued)

This table is wider than the reading column. Scroll horizontally to read all columns.

o2at table 20
Acronym or termMeaning in this document
Jitter / phase noiseTiming variation and its frequency-domain description. Clock quality must be stated with conditions and a measurement bandwidth.
Layer 2 / Layer 3Ethernet link-layer forwarding versus network-layer routing. Ordinary switches bridge o2at frames; IP routers do not route the raw protocol.
MAC / PCS / PMAMedia access control, physical coding sublayer and physical medium attachment. These layers provide framing, coding and serial functions around the optical module.
MADIMultichannel Audio Digital Interface, standardised by AES10. A dedicated professional serial audio interface.
MSA / SFFMulti-Source Agreement and Small Form Factor specifications. They standardise aspects of modules and interfaces, not a complete audio protocol.
MSB / LSBMost significant bit and least significant bit. Context distinguishes this numerical abbreviation from the manufacturer MSB Technology.
MTUMaximum transmission unit. The baseline Ethernet payload is at most 1,500 bytes, including the 12-byte o2at header.
Native DSDDSD transported without a DoP marker/container wrapper. It does not imply a universal USB implementation or automatic support by a DAC.
Nonce / sessionA fresh association value; a negotiated immutable audio context. The proposed nonce/session fields prevent routine stale-context confusion, not deliberate forgery.
NRZNon-return-to-zero signalling. A component’s NRZ rate rating must be related to the actual pulse-width requirements of the intended encoded stream.
o2atThe lowercase identifier for Open Optical Audio Transport in protocol magic, code and schemas. O2AT is the document’s display name.
OM3 / OM4 / OS2Optical-fibre categories. OM3/OM4 are multimode choices for the proposed SX baseline; OS2 is used with appropriate single-mode optics.
PAUSEIEEE 802.3x link-local MAC flow control. It is not a forwarded audio feedback message or a solution to unlimited fixed-clock drift.
PCM32Pulse-code modulation carried in a 32-bit signed container. The negotiated number of valid bits may be 16, 24 or 32.

Glossary of acronyms and terms (continued)

This table is wider than the reading column. Scroll horizontally to read all columns.

o2at table 21
Acronym or termMeaning in this document
PCP / QoS / VLANPriority code point, quality of service and virtual LAN. Their effective treatment depends on switch configuration and admission policy.
PDVPacket-delay variation. Changes arrival timing and FIFO occupancy; it need not set the DAC’s sample timing.
PLL / fractional synthesisPhase-locked loop and generation of rationally related frequencies. Device architecture, division ratios, spurs and loop bandwidth determine actual performance.
PMMA / POFPolymethyl methacrylate and plastic optical fibre, commonly used in consumer optical audio cables.
ppmParts per million, a relative frequency measure. One ppm means a ratio error of 10⁻⁶, not a phase-noise specification.
Pro ISL / Pro USBMSB’s proprietary optical interface and its USB-to-optical adapter. These are not o2at or ordinary Ethernet transports.
PTP / gPTPPrecision Time Protocol and its IEEE 802.1AS-related network-time profile. Frequency, phase and time synchronisation requirements must be distinguished.
Q32.32An unsigned 64-bit fixed-point representation with 32 integer and 32 fractional bits, used by the optional group-rate pacing hint.
SerDesSerialiser/deserialiser. A qualified high-speed receive path includes the relevant recovery and alignment functions, not simply ordinary differential GPIO.
SFP / SFP+Pluggable optical-module form-factor families. Compatible mechanics do not establish the module’s line rate, wavelength or carried protocol.
TLVType-length-value record. The o2at control format uses eight-byte record alignment and explicit critical-extension semantics.
TOSLINKToshiba-origin consumer optical interface family. This paper distinguishes its installed capabilities from the scope of all IEC 60958 extensions.
UAC2USB Audio Class 2.0. Class support alone does not prove native DSD handling, high-bandwidth endpoint implementation or the highest proposed o2at rates.
VCO / VCXOVoltage-controlled oscillator and voltage-controlled crystal oscillator. Clock-loop choices influence reference tracking and local phase-noise performance.
Wire utilisationThe fraction of link time occupied by a stream, including framing, preamble and gap. It is larger than the raw audio payload rate.

Bibliography and primary references

References were checked on 19 November 2026. Standards and manufacturer documentation establish factual boundaries; bandwidth, latency and buffer figures identified as calculations are derived in this paper. Public product information does not disclose proprietary wire formats. An edition cited here is a design reference, not a claim of certification.

R01. IEC. IEC 60958-1:2021, Digital audio interface, Part 1: General. Public scope and publication record.

https://webstore.iec.ch/en/publication/62829

R02. IEC. IEC 60958-5:2021, Digital audio interface, Part 5: Consumer application enhancement.

https://webstore.iec.ch/en/publication/59808

R03. Toshiba Electronic Devices & Storage Corporation. TOTX1950A(F), fibre optic transmitting module, Rev. 3.0, 15 April 2026, pp. 1–3.

https://toshiba.semicon-storage.com/info/docget.jsp?did=37194&prodName=TOTX1950A(F)

R04. Toray. RAYTELA plastic optical fibre, material and optical product specifications.

https://www.electronics.toray/en/products/raytela/

R05. Cliff Electronic Components. TOSLINK optical connectors, JIS F05, JEITA and CP-1201 compatibility descriptions.

https://www.cliffusa.com/products/optical/toslink.htm

R06. DoP project. DoP Open Standard, Version 1.1, Method for transferring DSD Audio over PCM Frames, §§2–3.

https://dsd-guide.com/sites/default/files/white-papers/DoP_openStandard_1v1.pdf

R07. NXP Semiconductors. UM11732, I²S bus specification, Rev. 3.0, 17 February 2022.

https://www.nxp.com/docs/en/user-manual/UM11732.pdf

R08. Audio Engineering Society. AES10, Serial multichannel audio digital interface (MADI), standards catalogue and scope.

https://aes.org/publications/standards-store/?id=17

R09. Audio Engineering Society. AES67-2023, High-performance streaming audio-over-IP interoperability.

https://aes.org/publications/standards-store/?id=140

R10. Audinate. What are the benefits of AES67? Dante transport and RTP-over-IP interoperability explanation.

https://support.getdante.com/hc/en-gb/articles/5508266895391-What-are-the-benefits-of-AES67

R11. Avnu Alliance. Milan Specification, consolidated AVB interoperability specification and supporting IEEE standards.

https://avnu.org/resource/milan-specification/

R12. HiFi Rose. RS130–RD160 SFP Connection Guide, manufacturer-authored connection instructions.

https://cdn.shopify.com/s/files/1/0631/7608/1655/files/SFP_Connection_Guide_EN_68.pdf?v=1736876712

R13. HiFi Rose. RS130, RD160 and RSA720 owner manuals, manufacturer downloads. Optical USB and Ethernet interfaces are documented separately.

https://hifirose.com/download

R14. MSB Technology. About MSB fibre connections, Pro ISL optical media, compatibility and non-Ethernet warning.

https://msbtechnology.com/support/about-msb-fiber-connections/

R15. MSB Technology. Digital inputs, Pro ISL clock return and error-handling description.

https://msbtechnology.com/dacs/digital-inputs/

R16. MSB Technology. Pro USB adapter, USB conversion, clocking and format specifications.

https://msbtechnology.com/dacs/prousb/

R17. Aurender. N50 Quick Start Guide, optional MSB Pro ISL output and format capabilities.

https://ask.aurender.com/hc/en-us/articles/35347051953047-N50-Quick-Start-Guide

R18. Coherent. FTLF8519P3xyL SFP transceiver data sheet, 1000BASE-SX and power below 500 mW.

https://www.coherent.com/resources/datasheet/networking/optical-transceivers/sfp/ftlf8519p3byl-transceiver-ds.pdf

R19. Cisco. 10GBASE SFP+ Modules Data Sheet, updated 6 October 2026, Table 8.

https://www.cisco.com/c/en/us/products/collateral/interfaces-modules/transceiver-modules/data_sheet_c78-455693.html

R20. SNIA SFF Committee. SFF-8472, Management Interface for SFP+, Rev. 12.5a, 16 January 2026, internal retimer/CDR indication.

https://members.snia.org/document/dl/25916

R21. SFF Committee. SFF-8431, Rev. 4.1 with addendum, archived SFP+ electrical interface specification, §§1 and 3. Successors include SFF-8418 and SFF-8419.

https://members.snia.org/document/dl/25891

R22. Xilinx. High-Speed Serial I/O Made Simple, running disparity and 8b/10b coding explanation.

https://download.amd.com/docnav/documents/aem/books-serialio.pdf

R23. IETF. RFC 9542, IANA Considerations and IETF Protocol and Documentation Usage for IEEE 802 Parameters, April 2024. Local experimental EtherTypes.

https://www.rfc-editor.org/rfc/rfc9542.html

R24. IEEE 802.3 Ethernet Working Group. Standards and public working-group materials; MAC, 1000BASE-X and 10GBASE-R are the normative layer foundations.

https://www.ieee802.org/3/

R25. Skyworks. Si5341/40 Rev. D data sheet, low-jitter clock generators with MultiSynth fractional synthesis.

https://www.skyworksinc.com/-/media/Skyworks/SL/documents/public/data-sheets/Si5341-40-D-DataSheet.pdf

R26. Texas Instruments. LMK04828 data sheet, dual-loop jitter cleaner and clock distribution device.

https://www.ti.com/lit/ds/symlink/lmk04828.pdf

R27. Microsoft. How to transfer data to USB isochronous endpoints, USB high-speed transfer intervals and payload limits.

https://learn.microsoft.com/en-us/windows-hardware/drivers/usbcon/transfer-data-to-isochronous-endpoints

R28. XMOS. lib_xud: USB device library, XM-012026-UG v4.0.1, 8 December 2025, §10 and build options.

https://www.xmos.com/documentation/XM-012026-UG/pdf/lib_xud_v4.0.1.pdf

R29. Aurender. MC20 Quick Start Guide, §§7.2–7.3, external reference clocks and asynchronous USB.

https://ask.aurender.com/hc/en-us/articles/25352309294871-MC20-Quick-Start-Guide

R30. LAiV. Harmony µDDC technical specifications, clock input and audio input/output formats.

https://www.laiv.audio/harmony-uddc-specs

R31. Sipeed. Tang Mega 138K Pro Dock documentation, FPGA, SFP+ cages, transceiver rates and toolchain.

https://en.wiki.sipeed.com/hardware/en/tang/tang-mega-138k/mega-138k-pro.html

R32. Gowin Semiconductor. Arora V FPGA product documentation, GW5AST SerDes capabilities.

https://www.gowinsemi.com/en/product/arora-v-fpga/

R33. Seaman, M. Protocol identification in 802 LANs, IEEE 802.1 Maintenance contribution, draft 10 April 2020, §1.3.1. Experimental EtherType deployment scope.

https://ieee802.org/1/files/public/docs2020/maint-seaman-protocol-identification-0420-v00.pdf

R34. AMD. Tri-Mode Ethernet MAC Product Guide PG051, v9.0, Pause Control Frames, 11 December 2024.

https://docs.amd.com/r/en-US/pg051-tri-mode-eth-mac/Pause-Control-Frames

R35. Merging Technologies. Introduction to RAVENNA Technology, official technical overview of IP media transport and clocking.

https://merging.atlassian.net/wiki/spaces/PUBLICDOC/pages/4819332/Introduction+to+RAVENNA+Technology.

R36. RME. Digiface USB, optical ADAT/S/P-DIF capabilities and S/MUX channel counts.

https://rme-audio.de/digiface-usb.html

R37. IEC. IEC 61937-1:2021, Digital audio: interface for non-linear PCM encoded audio bitstreams applying IEC 60958, Part 1: General, including March 2024 corrigendum.

https://webstore.iec.ch/en/publication/62828

Tunji Durodola (LinuxProphet) | pocketOne OÜ | tunji@pocketone.eu

Footnotes

  1. 1

    R01. IEC 60958-1:2021, general interface scope; R05. Cliff TOSLINK connector specifications. Full publication details and links are listed in the bibliography.

  2. 2

    R03. Toshiba TOTX1950A(F) data sheet; R04. Toray RAYTELA plastic optical fibre. Full publication details and links are listed in the bibliography.

  3. 3

    R03. Toshiba TOTX1950A(F) data sheet. Full publication details and links are listed in the bibliography.

  4. 4

    R01. IEC 60958-1:2021, general interface scope; R02. IEC 60958-5:2021, consumer extensions. Full publication details and links are listed in the bibliography.

  5. 5

    R06. DoP Open Standard v1.1. Full publication details and links are listed in the bibliography.

  6. 6

    R07. NXP UM11732, I²S specification. Full publication details and links are listed in the bibliography.

  7. 7

    R08. AES10, MADI interface scope; R09. AES67-2023, interoperability scope; R10. Audinate, Dante and AES67; R35. Merging, RAVENNA technical overview. Full publication details and links are listed in the bibliography.

  8. 8

    R11. Avnu, Milan Specification. Full publication details and links are listed in the bibliography.

  9. 9

    R12. HiFi Rose RS130–RD160 connection guide; R13. HiFi Rose product manuals. Full publication details and links are listed in the bibliography.

  10. 10

    R17. Aurender N50 Quick Start Guide. Full publication details and links are listed in the bibliography.

  11. 11

    R14. MSB fibre connections guide. Full publication details and links are listed in the bibliography.

  12. 12

    R15. MSB Pro ISL digital input description; R16. MSB Pro USB specifications. Full publication details and links are listed in the bibliography.

  13. 13

    R05. Cliff TOSLINK connector specifications. Full publication details and links are listed in the bibliography.

  14. 14

    R06. DoP Open Standard v1.1; R37. IEC 61937-1:2021, encoded audio carriage. Full publication details and links are listed in the bibliography.

  15. 15

    R36. RME Digiface USB, ADAT and S/MUX specifications. Full publication details and links are listed in the bibliography.

  16. 16

    R24. IEEE 802.3 Ethernet standards. Full publication details and links are listed in the bibliography.

  17. 17

    R18. Coherent FTLF8519P3xyL data sheet. Full publication details and links are listed in the bibliography.

  18. 18

    R18. Coherent FTLF8519P3xyL data sheet; R19. Cisco SFP+ data sheet, power specifications; R20. SFF-8472, optional internal retimer/CDR; R21. SFF-8431, host channels and interfaces. Full publication details and links are listed in the bibliography.

  19. 19

    R22. Xilinx, 8b/10b running disparity. Full publication details and links are listed in the bibliography.

  20. 20

    R23. RFC 9542, local experimental EtherTypes; R33. IEEE 802.1 contribution, experimental identification. Full publication details and links are listed in the bibliography.

  21. 21

    R25. Skyworks Si5341/40 Rev. D data sheet; R26. TI LMK04828 data sheet. Full publication details and links are listed in the bibliography.

  22. 22

    R34. AMD PG051, PAUSE control frames. Full publication details and links are listed in the bibliography.

  23. 23

    R31. Sipeed Tang Mega 138K Pro documentation; R32. Gowin Arora V transceiver documentation. Full publication details and links are listed in the bibliography.

  24. 24

    R27. Microsoft, USB isochronous endpoint limits. Full publication details and links are listed in the bibliography.

  25. 25

    R28. XMOS lib_xud v4.0.1, high-bandwidth limits. Full publication details and links are listed in the bibliography.

  26. 26

    R07. NXP UM11732, I²S specification. Full publication details and links are listed in the bibliography.

  27. 27

    R29. Aurender MC20 guide, USB and clock input. Full publication details and links are listed in the bibliography.

  28. 28

    R30. LAiV Harmony µDDC technical specifications. Full publication details and links are listed in the bibliography.

Source note: this is an HTML rendering, not a replacement for the editable source layout. Table captions, footnote references and heading anchors are preserved for reading and review. When line breaks or pagination matter, cite the corresponding matching PDF page.

Return to specification status and downloads