Nvidia unveils a new generation of server AI chips

At the annual GTC 2022 conference, Nvidia анонсировала several new chips and technologies designed to accelerate the speed of artificial intelligence algorithm computations.

The corporation представила architecture for next-generation Hopper GPUs and the H100 chip built on it, designed for machine-learning tasks.

The device is manufactured on a 4-nm process and contains 80 billion transistors. It is the company’s first GPU to support PCIe Gen5 and to use HBM3, delivering memory bandwidth of 3 TB/s.

The company stated that the H100 is three times faster than its predecessor A100 for FP16, FP32 and FP64 computations, and six times faster for 8-bit floating-point operations.

«For training giant transformers, the H100 proved nine times more productive, training in just a few days rather than weeks», said Paresh Hariya, Nvidia’s Senior Director of Product Management.

The H100 will go on sale in the third quarter of 2022.

Nvidia also представила the Grace CPU Superchip, based on the ARMv9 architecture. It comprises two Grace chips linked by NVLink, delivering data transfer speeds up to 900 GB/s.

Grace CPU Superchip contains 144 ARM cores and consumes around 500 W. The chip supports LPDDR5x memory, delivering up to 1 TB/s bandwidth.

Grace CPU Superchip will be available in the first half of 2023.

In addition to hardware, Nvidia updated a range of AI software services, including the Maxine SDK for audio and video enhancement and the Rivay toolkit for developing natural-language processing systems.

The company announced the creation of a new AI supercomputer, Eos. The installation will be equipped with 4,600 H100 GPUs, delivering 18.4 exaflops of performance. The system will be used solely for internal company research.

Earlier, under regulatory pressure from the US, the EU and the UK, Nvidia abandoned the purchase of ARM for $40 billion.

In the same month, the chipmaker reported a decline in revenue from selling GPUs used for mining cryptocurrencies.

In January, Meta announced the creation of the world’s largest AI supercomputer based on Nvidia processors.

Subscribe to ForkLog News on Telegram: ForkLog AI — all the AI news!