Huawei Atlas 950 SuperPoD
Overview
Performance Metrics
| Precision | Peak (EFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP16 | — | — | — |
| FP8 | 1.00 | 977 | — |
| FP4 | 2.00 | 1,953 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- Ascend 950DT
- Count
- 1024
- Memory per GPU
- 144 GB
- Bandwidth per GPU
- 4.0 TB/s
Power & Efficiency
- Aggregate Power
- —
- TDP per GPU
- —
- FP8 Efficiency
- —
- Aggregate Bandwidth
- 4096.0 TB/s
System Details
Interconnect
UnifiedBus 2.0 (Lingqu) all-optical, 1.72 PB/s aggregate scale-up, 64 x 1.68 TB/s bidirectional per cabinet
Notes
Huawei's 1,024-card Atlas 950 SuperPoD as shown at WAIC 2026: 16 compute cabinets plus 4 UnifiedBus interconnect cabinets (44OU each), up to 1,024 Ascend 950DT, 1,024 x 96 GB on-chip memory at 4.0 TB/s, 1 EFLOPS mxFP8/FP8/HiF8 and 2 EFLOPS mxFP4, 1.72 PB/s total scale-up bandwidth, 256 TB globally addressable memory, fully liquid cooled. Huawei's HC2025 keynote of 18 September 2025 described an 8,192-card configuration of the same product delivering 8 EFLOPS FP8 and 16 EFLOPS FP4. Huawei states no dense or sparse basis for any of these figures. POWER: Huawei publishes no power rating for this machine at either the 1,024-card or the 8,192-card scale. Its only first-party page for the 1,024-card unit (huawei.com/cn/news/2026/7/atlas-950-superpod) carries no watt line at all, the e.huawei.com SuperPoD catalogue does not list the Atlas 950 yet, and Huawei has never published a chip-level power figure for any Ascend part. The 100 kW previously stored here cannot be a whole-SuperPoD rating, since 100 kW over 1,024 cards is 98 W per card; a per-cabinet reading would be arithmetically plausible against the 16 compute cabinets, but no Huawei source has been found for it at any scope, so nothing is stored rather than a rescaled guess.