Google unveiled its new option for companies wanting a custom AI supercomputer in the cloud at its I/0 conference. The A3 supercomputer consists of new virtual machines and is equipped with the most powerful GPUs from Nvidia.

Want more AI? Then you need more computing power. This is also recognized by Google, which is presenting the A3 supercomputer at its I/O conference. These are HPC clusters running on the Google Cloud Platform and made up of powerful building blocks: the A3 instances.

Google uses an architecture reminiscent of the blueprint used by competitor Microsoft for its supercomputer Azure. Microsoft has connected tens of thousands of Nvidia GPUs together using a newly developed connection system. Google is also using a new system to link GPUs together. This is based on a brand new Infrastructure Processing Unit (IPU). This allows GPUs to communicate with each other at 200 Gbps without taxing the CPU.

specifications

The new A3 instances also feature fourth-generation Intel Xeon Scalable processors paired with eight Nvidia Hopper H100 GPUs. NVSwitch and NVLink 4.0 are also present. Finally, each host gets 2TB of DDR5 memory.

Google offers an end-to-end solution with the A3 instances combined with Vertex AI. Of course it is also possible to bring your own software stack or to work with Kubernetes. The A3 supercomputer instances are suitable for training new (generative) AI models, but also for inference.

Good for Nvidia

HPC instances are gradually becoming a must for cloud providers. This is very good news for Nvidia, which currently enjoys a virtual monopoly when it comes to powerful AI accelerators. Finally, the Nvidia H100 has no equal and is therefore the logical choice for any cloud provider looking to build an AI supercomputer. The production lines are therefore running at full speed and Jensen Huang’s till is ringing.