Intel Builds Large AI Supercomputer: Xeon Meets Gaudi2

Intel on Tuesday announced one of the world’s first large deployments of its Gaudi2 accelerators for artificial intelligence (AI) workloads. The supercomputer — built in collaboration with Dell — will be used by Stability AI for its generative AI applications.

Intel’s new supercomputer is based on the company’s latest Intel Xeon Scalable processors and runs as many as 4,000 Intel Habana Gaudi2 AI hardware accelerators. Assuming that each machine employs eight Habana Gaudi2 processors in OAM form-factor (which is typical for OAM machines), the machine features as many as 500 nodes. Given that Intel’s Gaudi2 offers formidable performance that’s comparable to that of Nvidia’s H100 in select applications, the supercomputer should be quite powerful. We’re probably looking at 7–8 FP16 ExaFLOPS of AI performance.

Intel and Dell did not disclose whether they will continue to team up for AI supercomputers going forward. Meanwhile, for the chip designer, it’s important that one of the world’s largest server suppliers has PowerEdge machines with its Xeon Scalable CPUs and Habana Gaudi2 accelerators developed specifically for AI workloads. These machines are designed to handle both large-scale training and base-level inferencing.

Besides Dell, Alibaba Cloud said that the 4th Generation Xeon can achieve a 3x inference speed in its model-serving platform DashScope program when used with Advanced Matrix Extensions (Intel AMX) accelerators and various software optimizations.

Intel’s processors power numerous supercomputers and even Nvidia’s DGX H100 systems for AI and high-performance computing (HPC) applications, but the company’s own AI accelerators and compute GPUs are not as lucky and there are very few Gaudi/Gaudi2 or Ponte Vecchio-based AI deployments to date. With Stability AI working with Gaudi2, it will be interesting to see what other companies also choose to give it a shot.