Intel announced its latest chip to challenge Nvidia’s dominant market position for semiconductors that train and deploy large AI models. The company claims its Gaudi 3 chip is projected to deliver an average 50% improvement in the time taken to train Meta Platform’s open-source Llama2 models compared with Nvidia’s H100 processor launched in 2023.
Intel predicted its Gaudi 3 accelerator inference throughput would typically outperform the H100 chips for some of the Llama and Falcon models it tested by up to 50% and stated its chips use less power than Nvidia silicon. It stated the Gaudi 3 AI accelerator would power systems with tens of thousands of accelerators connected through ethernet and said the chip would deliver a significant leap in AI training and inference for global enterprises looking to deploy GenAI at scale.
Intel employed Taiwan Semiconductor Manufacturing Co’s 5nm process to build the chips. The chip will be widely available in Q3 to Dell, HPE, and Supermicro, among others. Last month, Nvidia announced its flagship B200 Blackwell chip, which can perform certain tasks 30 times faster than the H100. Nvidia has an estimated 80% share of the AI chip market, which has propelled it to new heights over the past year.
IBM, in collaboration with the US Department of Energy-funded Quantum Science Center (QSC) at Oak Ridge National Laboratory (ORNL) and AMD, has introduced a co-design platform for quantum error correction that integrates quantum and high-performance computing.
IBM, in collaboration with Siemens, has announced new research intended to advance nanostack architecture, a novel transistor framework introduced earlier this year.