Atlas 950 SuperPoD product page · inference-research

Fields

id

sources/cn-huawei-atlas-950-superpod-spec

URL

https://www.hiascend.com/en/hardware/cluster?tag=950

Publisher

Huawei HiSilicon / Ascend Community

Kind

spec-sheet

Accessed

2026-10-03

notes

PRIMARY VENDOR SOURCE. 44OU compute cabinet with 64x 950DT NPU: 64x 96 GB on-chip memory, 4.0 TB/s, 114.1 PFLOPS mxFP4 / 58.8 PFLOPS mxFP8-HiF8 / 31.1 PFLOPS FP16-BF16, 64 x 1.68 TB/s bidirectional UB Link. Scales to 1024 NPUs at 1.72 PB/s UB Link with 256 TB unified global memory addressing, ~100 kW liquid cooled. The 256 TB unified global memory claim is the single most consequential number here for anyone modelling Ascend memory semantics.

Notes

PRIMARY VENDOR SOURCE. 44OU compute cabinet with 64x 950DT NPU: 64x 96 GB on-chip memory, 4.0 TB/s, 114.1 PFLOPS mxFP4 / 58.8 PFLOPS mxFP8-HiF8 / 31.1 PFLOPS FP16-BF16, 64 x 1.68 TB/s bidirectional UB Link. Scales to 1024 NPUs at 1.72 PB/s UB Link with 256 TB unified global memory addressing, ~100 kW liquid cooled. The 256 TB unified global memory claim is the single most consequential number here for anyone modelling Ascend memory semantics.

Cross-references

This record references no other records.