NVIDIA claims the Ampere-based A100 has 20x more AI performance than Volta and 2.5x higher double-precision performance. It also supports FP64 numbers.
NVIDIA claims that the GPU will be 20x more powerful than Volta GPU in AI-training (single-precision operations) and AI interference (8-bit integer operations). In terms of high-performance computing (double precision operations) NVIDIA says that the GPU will be 2.5 faster.
“With this new precision, A100 offers 20 times more compute for single-precision AI, and because developers can continue to use the inputs as single-precision and get outputs back as single-precision, they do not need to do anything differently. They benefit from this acceleration automatically out of the box,”
NVIDIA Kharya, director of product management for accelerated computing at Nvidia
Later today, NVIDIA is expected to showcase its DGX A100. This system will pack eight Tesla A100 GPUs to offer up to 5 petaflops of computing power. While the previous generation was primarily pitched for AI-operations, the DGX A100 can also handle scale-up operations (data analytics- and scale-out operations (inferencing). For at least some computing tasks, a DGX A100 rack can replace 25 CPU datacenter racks, at 1/10th the investment and 1/20th the power consumption.