Google Launches AI Supercomputer Powered by Nvidia H100 GPUs
Google launches Google I/O We spent over an hour discussing the many advances in artificial intelligence this afternoon. The company described its new PaLM 2 large-scale language model (LLM) for generative AI that powers its Bard chatbot tool. This is the foundational pillar for adding AI-infused capabilities across Google’s product portfolio, including Google Maps, Google Photos, and Gmail (among others).
With that in mind, millions (and ultimately billions) of users submit requests for operations as mundane as removing a person remaining in the background of a photo. So, to power the real-world model, we need a lot of computing power in the cloud. Compose an entire email based on a short text prompt. That’s where Google’s new A3 GPU supercomputer comes in. According to Google, the new A3 supercomputer is “purpose-built to train and serve the most demanding AI models, driving innovation in today’s generative AI and large-scale language models,” with 26 exaflops of AI. performance is achieved.
Each A3 supercomputer is powered by 4th Gen Intel Xeon Scalable processors with 2TB of DDR5-4800 memory. But the real “brain” of this operation comes from eight of his Nvidia H100 “Hopper” GPUs, which leverage NVLink 4.0 and NVSwitch to access 3.6 TBps of bisection bandwidth.
According to Google, the A3 is the first production-level deployment of GPU-to-GPU data interfaces, allowing data sharing at 200 Gbps while bypassing the host CPU. This interface, which Google calls the Infrastructure Processing Unit (IPU), increases available network bandwidth for A3 virtual machines (VMs) by a factor of 10 compared to A2 VMs.
“Google Cloud’s A3 VMs, powered by next-generation NVIDIA H100 GPUs, accelerate the training and delivery of generative AI applications,” said Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA. “Google Cloud’s recent release of he G2 instances continues with our continued work with Google Cloud, and our dedicated He is proud to help transform enterprises around the world with AI infrastructure.”
If your business wants to leverage A3 virtual machines, the only way to get access is through Google’s A3 preview interest form To join the early access program. However, as Google clearly states, entering information does not guarantee participation in the program.