In a rapid move, Google Cloud has made the NVIDIA A100 Tensor Core GPU a part of their cloud offerings, just weeks after the launch date. The new data center-focused A100 was designed with AI training and inference, data analytics, scientific computing, genomics, and edge video analytics in mind.
The A100, based on the new Ampere architecture, offers significant improvements in compute performance as compared to its predecessor, the V100. Artificial intelligence model training and inference computing performance can be boosted up to 20 times as compared to the previous generation GPU.
A100 GPUs can be found in the Accelerator-Optimized VM (A2) family of Virtual Machine (VM) instances, with up to a whopping 16 GPUs available to a single instance. This configuration provides 640 GB of GPU memory and an effective performance of up to 10 petaflops of FP16 data, or 20 petaops of int8 data. For those with more modest compute needs, VMs with a single GPU are also available.
At present these instances are available through the private alpha program, with public availability scheduled for later this year.