NVIDIA H800 80GB SXM (China export SKU)

NULL, and that is the honest state. No vendor or secondary source reached for this record publishes an H800 memory bandwidth figure. NVIDIA publishes no H800 datasheet at all, so there is no primary number to record and none is inferred from the H100's 3.35 TB/s.

RECORD EXISTS TO KEEP A WARNING, NOT TO SUPPLY NUMBERS. This accelerator has almost no trustworthy specification, and the honest thing is to say so on the record rather than to fill it with plausible H100 numbers. NVIDIA has NEVER published an H800 datasheet. On record, NVIDIA declined even to describe the part: a spokesperson 'declined to say how the China-focused H800 differs from the H100, except that "our 800 series products are fully compliant with export control regulations"' (acc-fill-reuters-h800-export-china). That single refusal is why every numeric field on this record is null. WHAT IS PUBLISHED - the reduced interconnect, as a QUALITATIVE claim, not a number. Two independent sources agree that chip-to-chip transfer was cut, and neither gives a GB/s figure they sourced from NVIDIA. Reuters, via an unnamed 'chip industry source in China': 'the H800 mainly reduced the chip-to-chip data transfer rate to about half the rate of the flagship H100' (acc-fill-reuters-h800-export-china). Data Center Dynamics, attributing the same Reuters reporting: 'cutting its interconnect performance by about half - leaving it with around 300GBps' (acc-fill-dcd-h800-interconnect-halved). The 'around 300 GB/s' figure is DCD's gloss on the same claim, and since a Hopper SXM is normally quoted at 900 GB/s it is a third, not a half - so the two phrasings are loose approximations of one qualitative fact, not two measured values. The interconnect list above records that this part HAS NVLink 4 and NVSwitch 4 (it ships in 8-GPU NVLink+NVSwitch nodes, per DeepSeek's published H800 cluster description) WITHOUT recording a bandwidth, because no source I could reach states one for the H800. Downstream bundles that quote 'H800 = 400 GB/s' are repeating an unattributed number; nothing here endorses it. WHAT IS PUBLISHED - reduced FP64: NOT SUBSTANTIATED, and this is the deliberate null. The task framing for this record assumed H800 had a documented FP64 reduction, and I could not find a single source that says so. Searches across multiple engines and the accessible trade press turned up reporting that describes the cut as INTERCONNECT-ONLY: CNBC on the October 2023 rule change writes that buyers 'were able to buy a slightly slowed-down version called the H800 or A800 that complies with U.S. restrictions, primarily by slowing down an on-device connection speed, called an interconnect' (acc-fill-cnbc-h800-export-ban) - the 'primarily' is not upgraded to 'exclusively' there either, but no source goes further than that. So: the FP64 row is not recorded, and flops is an empty list rather than a copy of the H100's figures. A reader who wants H800 FP64 must treat it as unknown. This is exactly the 'H800 is just an H100' error the parent task flagged, and the disciplined answer cuts both ways: it is ALSO an error to assert a specific FP64 cut figure when no source prints one. WHY contested AND NOT verified: the export-control framing and the existence of the part are confirmed by on-record sources, but not one performance figure is vendor-confirmed. Per AGENTS.md a single secondary source gets contested, and here even the interconnect claim rests on an unnamed industry source relayed by wire reporters. Confidence 0.4 reflects a part whose existence is certain and whose specification is unpublished. OTHER FIELDS LEFT NULL AND WHY: vram_gb and memory_type are null because no source states the H800's capacity or memory generation, even though '80GB' is widely assumed by analogy to H100 SXM - the 80 GB is in this record's NAME because that is how the part is universally referred to and how the citing supply record labels it, but it is not a figure any source I reached prints, so it is not recorded as a field value. process_nm is null (no source states it; do not inherit 4 nm from Hopper). tdp_w is null. unified_memory is null. form_factors 'sxm' reflects how it ships (8-GPU SXM modules in NVSwitch nodes), which DeepSeek's cluster description supports, but the physical module is not described by NVIDIA. SUPPLY-CONTEXT, which is why a used-market record for this part is reasonable: the October 2023 US rule change closed the loophole and 'the new export restrictions will restrict the export of Nvidia's A800 and H800 chips' (acc-fill-cnbc-h800-export-ban). So new H800 supply ended in 2023 and any H800 seen today is secondary-market or pre-rule inventory. Historically important because H800 is the chip DeepSeek-V2 and V3 were trained on (2.788M H800 GPU hours for V3), so it is load-bearing for the ML-cost narrative in this repo - and the interconnect reduction is the concrete reason a cluster of H800s was measurably more expensive per unit of trained-token throughput than a cluster of H100s would have been.