Article

Intel has launched the 'Habana Gaudi4,' a custom AI accelerator aimed at speeding up large-scale model training and inference. The Gaudi4 includes optimized memory paths and scalable multi-chip interconnects, catering to hyperscalers and enterprise AI teams.

It enhances previous Habana designs with increased throughput, improved tensor cores, and support for major ML frameworks. The chips integrate with Intel’s data center stack, and partnerships with cloud providers are established for early access and benchmarking of transformer workloads. Gaudi4 aims to lower infrastructure costs for organizations operating large language models and highlights a shift towards diverse AI hardware solutions.