This is for Google customers
Google has unveiled its new A3 supercomputers, which are specifically designed to train and serve the most demanding AI models that underpin today’s generative AI and innovation in large language models. Interestingly, these supercomputers are based on Nvidia H100 accelerators, although Google has its own designs for such tasks.
Apparently, this is due to the fact that the H100 is the latest and very productive solution for which Google has no alternatives yet.
The press release says the A3 is the first to use purpose-built 200 Gb/s interfaces, with data transfer from one GPU to another bypassing the CPU and passing through separate interfaces from other VM networks and data traffic. This provides up to 10 times the network bandwidth compared to A2.
Google unveils A3 supercomputer based on Nvidia H100 GPU
Google also notes that its Jupiter data center intelligent network fabric scales to tens of thousands of highly connected GPUs and allows full bandwidth reconfigurable optical links that can adjust the topology on demand.
The scalability of the A3 allows Google to talk about 26 exaflops of AI performance.
Apparently, under the name A3 Google does not mean some conditional supercomputer, but a very specific system consisting of eight H100 GPUs, an Intel Xeon Sapphire Rapids processor and 2 TB of DDR5-4800 RAM.
It’s worth noting that A3s are for Google customers, not the company itself.