IBM has announced plans to make Intel's Gaudi 3 AI processor available in its public cloud platform.

The Gaudi 3 chip is based on TSMC's five-nanometer node and features two types of computing modules: MMEs for matrix multiplications and TPCs for other AI-related calculations. The chip includes 64 TPCs and 4x as many MMEs as its predecessor, supported by a 120 GB memory pool with reportedly higher clock speeds. IBM plans to offer the chip in its IBM Cloud Virtual Servers for VPC early next year.

Intel claims that Gaudi 3 can perform inference with up to 2.3x the power efficiency of NVIDIA's H100 while training some large language models (LLMs) can be trained in less time. The chip also features an onboard Ethernet module for linking processors and servers, with a doubled bandwidth of 200 Gbps for individual Ethernet networking links.

Analyst QuickTake : Since the launch of its Gaudi3 AI chip in December 2023 , Intel has been attempting to challenge the dominance of NVIDIA in the AI chip space with its improved efficiency, where the latter has faced some production delays due to a design flaw . Despite this, Intel faces an uphill struggle where NVIDIA is expected to ramp up its production of its Blackwell series of chips later in the year, offering 25x lower cost and energy consumption than its predecessors.