Chip maker Nvidia has unveiled a brand new AI chip constructed for accelerated computing and to deal with the world’s most advanced generative AI workloads, spanning giant language fashions, recommender techniques, and vector databases.
The subsequent-generation GH200 Grace Hopper platform relies on a brand new Grace Hopper Superchip with the world’s first HBM3e processor, which might be obtainable in a variety of configurations, the corporate stated.
“The brand new GH200 Grace Hopper Superchip platform delivers this with distinctive reminiscence expertise and bandwidth to enhance throughput, the power to attach GPUs to mixture efficiency with out compromise, and a server design that may be simply deployed throughout all the knowledge centre,” Jensen Huang, founder and CEO of Nvidia, stated in a press release.
The brand new platform makes use of the Grace Hopper Superchip, which could be linked with further Superchips by Nvidia NVLink, permitting them to work collectively to deploy the enormous fashions used for generative AI.
This high-speed, coherent expertise offers the GPU full entry to the CPU reminiscence, offering a mixed 1.2TB of quick reminiscence when in twin configuration, based on the corporate. “HBM3e reminiscence, which is 50 per cent quicker than present HBM3, delivers a complete of 10TB/sec of mixed bandwidth, permitting the brand new platform to run fashions 3.5x bigger than the earlier model, whereas enhancing efficiency with 3x quicker reminiscence bandwidth,” Nvidia stated.
Main system producers are anticipated to ship techniques primarily based on the platform in Q2 of calendar yr 2024.