Huawei Atlas 800I A3
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP16 | 4.5 | 560.0 | — |
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
- 14.6 kW
- TDP per GPU
- 600 W
- FP16 Efficiency
- 0.31 TFLOPS/W
- Aggregate Bandwidth
- 25.6 TB/s
System Details
Interconnect
UnifiedBus (UB): 784 GB/s bidirectional D2D per NPU; 8 x 400GE QSFP-DD direct from NPU over RoCE, plus 56 x 400GE QSFP-DD over the bus protocol
Notes
The inference member of Huawei's Atlas 800 A3 family: a 10U air-cooled rack server with eight Ascend 910 NPUs and four Kunpeng 920 CPUs, rated at 4.48 PFLOPS of FP16 with 1,024 GB of on-chip memory across the eight NPUs at 3.2 TB/s each. That works out at 560 TFLOPS per NPU against the 750 of the training-oriented Atlas 800T A3 in the same chassis, which is a genuine inference bin rather than a transcription difference: Huawei's own family page lists three Atlas 800 A3 models at 6.0, 5.0 and 4.48 PFLOPS FP16. Memory is offered as 8 x 128 GB or 8 x 64 GB; the 128 GB configuration is recorded here. Networking is eight 400GE QSFP-DD ports direct from the NPUs over RoCE plus fifty-six more over Huawei's bus protocol, with 784 GB/s of bidirectional device-to-device bandwidth. Power is up to 14.6 kW from six hot-swap 3.0 kW supplies in 5+1 redundancy, air cooled. Huawei publishes no INT8 figure for this machine and labels nothing dense or sparse, so only the FP16 rate it states is recorded and nothing is derived.