Huawei Atlas 800T A3
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP16 | 6.0 | 750.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- Ascend 910C
- Count
- 8
- Memory per GPU
- 128 GB
- Bandwidth per GPU
- 3.2 TB/s
Power & Efficiency
- Aggregate Power
- —
- TDP per GPU
- 600 W
- FP16 Efficiency
- —
- Aggregate Bandwidth
- 25.6 TB/s
System Details
Interconnect
UnifiedBus (UB): 784 GB/s bidirectional D2D per NPU; scales to a 384-NPU Atlas 900 A3 SuperPoD
Notes
The training member of Huawei's Atlas 800 A3 family: a 10U air-cooled compute node with eight Ascend 910 NPUs and four Kunpeng 920 CPUs, rated at 6.0 PFLOPS of FP16 with 1,024 GB of on-chip memory at 3.2 TB/s per NPU and 784 GB/s of bidirectional device-to-device bandwidth. Huawei sells three models on this chassis at 6.0, 5.0 and 4.48 PFLOPS FP16, which works out at 750, 625 and 560 TFLOPS per NPU; this is the top one, and the 4.48 model is the separately listed Atlas 800I A3. Multiple nodes combine into an Atlas 900 A3 SuperPoD of up to 384 cards. Huawei publishes only an FP16 figure for this machine, and no INT8: a previously held INT8 value of 12.0 PFLOPS was 8 x 1,500 TFLOPS, a per-NPU rate Huawei states nowhere, and has been removed. Power is also not attributable: Huawei's family page gives 16.2 kW as the maximum input power for the Atlas 800 A3 line, but that line covers three compute variants and the Atlas 800I A3's own page separately says 14.6 kW, so no figure is recorded here rather than guessing which model the 16.2 kW describes. Huawei labels nothing dense or sparse.