NVIDIA unveils the A100 GPU in PCIe form

NVIDIA has unveiled its latest GPU based on Ampere technology. However, it is not a completely new card, but one with A100 chips again, albeit PCIe-enabled this time. This, then, is another potentially valuable addition to this growing line-up of cards for high-demand computing that may involve AI or machine learning.

The OEM has already technically launched the A100 series, introducing it along with its new Tensor Float 32 (TF32) and floating-point 64 (FP64) formats for the same. NVIDIA claims they improves the speeds at which AI and HPC tasks are performed by 20 and 2.5 times respectively compared to Volta-based GPUs.

The chip-maker has also announced that some major manufacturers have designed servers based on the new PCIe-based A100 GPUs. They include Supermicro; Lenovo; Hewlett Packard Enterprise; Asus and Gigabyte. These OEMs will soon supply hardware that can support up to 10 of these new cards in some cases.

Meanwhile, other companies, Dell, Cisco and Fujitsu included, have signed up to make servers that incorporate the third-gen Tensor Core versions of A100 GPUs.

Either kind of server has also now been issued with Ampere-compatible software, including an NVIDIA SDK for their take on HPC; Jarvis, an AI-services framework geared towards conversational use-cases; Merlin, a different framework for recommendations-based applications; the RAPIDS open-source data-science software library suite and CUDA 11.