Huawei Atlas 650E
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP16 | 3.4 | 425.0 | — |
| BF16 | 3.4 | 425.0 | — |
| FP8 | 6.4 | 803.8 | — |
| FP4 | 12.5 | 1560.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- Ascend 950DT
- Count
- 8
- Memory per GPU
- 144 GB
- Bandwidth per GPU
- 4.0 TB/s
Power & Efficiency
- Aggregate Power
- 14.5 kW
- TDP per GPU
- —
- FP8 Efficiency
- 0.44 TFLOPS/W
- Aggregate Bandwidth
- 32.0 TB/s
System Details
Interconnect
UnifiedBus (Lingqu): 8 x 784 GB/s bidirectional within a node, 2 x 8 x 1.68 TB/s across a two-node pair; UBoE direct from NPU and RoCE both at 400 Gbps per NPU
Notes
A 14U air-cooled rack server with eight Ascend 950DT NPUs and two Kunpeng 950 CPUs, built for ordinary air-cooled machine rooms rather than liquid-cooled halls. Huawei rates it at 12.48 PFLOPS mxFP4, 6.43 PFLOPS across mxFP8, FP8 and HiF8, and 3.40 PFLOPS FP16/BF16, with 768 GB of on-chip HBM across the eight NPUs at up to 4.0 TB/s each. Two machines can be joined over UnifiedBus into a sixteen-NPU full-mesh. Power is up to 14.5 kW from six hot-swap 3.0 kW supplies in 5+1 redundancy, cooled by thirty hot-swap fan modules. Note that the per-NPU figure works out at 1.56 PFLOPS mxFP4, below the 1.95 PFLOPS the same Ascend 950DT reaches in the liquid-cooled Atlas 950 SuperPoD: this is a de-rated air-cooled bin, not a transcription error, and should not be reconciled upward. Huawei labels none of these figures dense or sparse and publishes no sparse rate; its asterisk on the compute row means only that the values are theoretical and measured results may differ by under one percent.