Huawei Atlas 950 SuperPoD product page - 64-NPU AI Performance and on-chip memory table, an INDEPENDENT second server figure for the same 950DT

VERBATIM 2026-10-04 via web_extract. Recorded because it gives Huawei's AI Performance for a DIFFERENT NPU count (64 rather than 8) on the same 950DT silicon, which is the only way to test whether the stored per-NPU figures are right without assuming all NPUs in a cabinet perform identically. 'Atlas 950 SuperPoD Product Specifications' TABLE, verbatim: 'Form Factor | 44OU compute cabinet ≈ 47U compute cabinet' 'NPU | 64 × 950DT NPU' 'CPU | 16 × 950 CPU (up to 96 cores, 2.3GHz)' 'On-chip Memory | 64 × 96 GB total, 4.0 TB/s peak memory bandwidth' 'AI Performance | 114.1 PFLOPS @ mxFP4 58.8 PFLOPS @ mxFP8/HiF8 31.1 PFLOPS @ FP16/BF16' 'UB Link Bandwidth | Support up to 64 x 1.68 TB/s (bidirectional)' 'Power Usage | ～100 kW' 'Weight | 1,700 kg' 'Dimensions | 2,250 mm × 600 mm × 1,400 mm (excluding front & back doors)' Product Features prose, verbatim: 'The largest SuperPoD, supporting up to 1,024 NPUs with 1.72 PB/s UB Link bandwidth', 'Support 256 TB unified global memory addressing & memory semantics', 'Support up to 0.9 EFLOPS mxFP8/FP8/HiF8 compute power', 'Supporting liquid cooling with 100 kW power supply'. DISCLAIMER, VERBATIM: '*Theoretical computing power may have a <1% error in actual testing.' and '*The specific configuration is subject to the actual configuration.' CROSS-CHECK AGAINST THE ATLAS 650E (8-NPU) PAGE, AND THE RESULT IS A REAL DISCREPANCY IN HUAWEI'S OWN FIGURES. Dividing this 64-NPU cabinet's figures by 64 gives per-NPU 1.783 PFLOPS mxFP4, 0.919 PFLOPS mxFP8/HiF8, 0.486 PFLOPS FP16/BF16. Dividing the Atlas 650E's 8-NPU figures by 8 gives 1.56, 0.804, 0.425. So the two Huawei pages imply PER-NPU FIGURES THAT DIFFER BY 14.3%, 14.3% and 14.3% RESPECTIVELY - consistently, across all three precisions. This is not rounding noise. THE MOST LIKELY EXPLANATION, STATED AS INFERENCE NOT FACT: the 8-NPU Atlas 650E is a single air-cooled server while the 64-NPU Atlas 950 SuperPoD is a liquid-cooled cabinet rated at ~100 kW, so the two products plausibly run the same 950DT silicon at different clocks and power envelopes. Huawei's own page names this as the reason in spirit by listing 'Supporting liquid cooling' as a feature. But NO Huawei document states a per-NPU figure or a clock for either product, so this cannot be settled from vendor material and the two per-NPU derivations must both be retained. CONSEQUENCE FOR accelerators/huawei-ascend-950dt: its stored 1,560 / 803.75 / 425 TFLOPS per-NPU values are derived correctly from the Atlas 650E (8-NPU) page and are defensible AS THAT. They are NOT the values implied by Huawei's own 64-NPU SuperPoD page, which implies 1,783 / 918.75 / 486 TFLOPS per NPU. The record should either state which server its per-NPU figures describe (it currently describes 'Atlas 650E / 950 SuperPoD' as if the two agreed) or carry both derivations. Its memory_bandwidth_basis already names the Atlas 650E specifically, which is correct; its flops basis_detail does not.