| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP16 | 307.2 | 800.0 | — |
| BF16 | 307.2 | 800.0 | — |
| INT8 | 614.4 | 1600.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
UnifiedBus (UB) all-to-all optical; 784 GB/s bidirectional D2D per NPU
Huawei's 384-NPU supernode, built from twelve 47U liquid-cooled compute cabinets plus four air-cooled bus-equipment cabinets, each 2,250 by 600 by 1,150 mm. It carries up to 384 Ascend 910 NPUs and 192 Kunpeng 920 CPUs, with 384 x 128 GB of on-chip memory at up to 3.2 TB/s each and 784 GB/s of bidirectional device-to-device bandwidth. Huawei rates it at up to 307.2 PFLOPS of FP16, and also lists 288.7 and 240.3 PFLOPS variants; the top figure works out at exactly 800 TFLOPS per NPU, which is the Ascend 910C's own published rate, so the BF16 and INT8 figures here are that chip's published 800 and 1,600 TFLOPS multiplied by 384. Cooling is liquid for the compute cabinets and air for the bus cabinets, with an operating range of 5 to 40 C. Note that Huawei publishes no power figure for this machine: its supply row states only the input voltage, three-phase 380V AC on two feeds. A figure of 559 kW previously held here appeared in no Huawei source and has been removed rather than shipped. Huawei labels nothing dense or sparse. This machine is also sold as CloudMatrix 384, which is a deployment name for the same 384-NPU configuration rather than a separate product.