NVIDIA A800 SXM4

The NVIDIA A800 SXM4 is the mezzanine form of the A800, the export-compliance version of the A100 that NVIDIA sold into China, Hong Kong and Macau from late 2022. It is the same GA100 silicon in the same configuration as an A100 80GB SXM, and the compute figures are identical to the decimal place: 9.7 TFLOPS of FP64, 19.5 of FP32, 156 of TF32 and 312 of FP16 and BF16, with 80 GB of HBM2e at 2,039 GB/s and a 400 W board power. One thing changes, and it is the thing the export rules of the day measured. NVLink runs at 400 GB/s per GPU rather than the A100's 600 GB/s, which slows the all-to-all traffic that large-model training depends on while leaving single-GPU throughput untouched. That makes the A800 an unusually clean illustration of what an interconnect ceiling costs: identical silicon, identical arithmetic, a third less bandwidth between chips. Eight of these modules sit on an HGX A800 baseboard, where NVSwitch gives any pair of GPUs a direct 400 GB/s path, which is how they reach the market in machines such as Inspur's NF5688M6. The part had a short life. The United States tightened the rules again in October 2023, adding the A800 itself to the controlled list, and OEM product guides for it are published today as withdrawn.