IBM Cloud to Introduce Intel Gaudi 3 AI Processor
IBM recently announced plans to introduce Intel's Gaudi 3 artificial intelligence processor on its public cloud platform. This processor is considered an alternative option to Nvidia's popular H100 graphics processing unit (GPU). The H100 has been Nvidia's flagship AI accelerator until the release of the faster Blackwell B200 GPU, which is expected later this year.
According to Intel, the Gaudi 3 processor offers up to 2.3 times better energy efficiency and reduced training time for large language models compared to the H100. Gaudi 3 is the third-generation product in Intel's AI processor series, which was acquired through the purchase of a startup company in 2019. It is manufactured using TSMC's 5-nanometer process.
The processor provides processing power through two sets of computing modules: MMEs and TPCs, each optimized for different tasks. The MME module focuses on executing matrix multiplication operations, which are crucial mathematical calculations for converting input data into decisions in AI models. The TPC module, on the other hand, is based on a Very Long Instruction Word (VLIW) architecture and aims to parallel process multiple computations to accelerate AI model performance. Gaudi 3 has 64 TPCs, nearly three times more than its predecessor, and the number of MMEs has quadrupled.
In addition to performance improvements, Gaudi 3 is equipped with a 120GB high-speed memory pool, which has a faster RAM clock speed compared to Intel's previous generation AI processor. It can achieve a peak speed of 18.35 million trillion floating-point operations per second (TFLOPS) when processing BF16 data types, which are commonly used in AI models.
Apart from the performance enhancements, Gaudi 3 also features a built-in Ethernet module that supports interconnectivity between AI servers, with the bandwidth of a single Ethernet link doubled to 200 gigabits per second.
IBM plans to offer Gaudi 3 services on its IBM Cloud Virtual Servers for VPC public cloud platform in early next year. It will also add support for Gaudi 3 in its WatsonX product suite, which includes software tools for building, deploying, and executing AI models and related tasks.