On Tuesday, Nvidia revealed a newly-created chip tailored to executing artificial intelligence models. The GH200 comes with the same GPU as the H100, the firm's premier AI chip, coupled with 141 gigabytes of sophisticated memory and a 72-core ARM CPU. Jensen Huang, the CEO of Nvidia, asserted that “This processor is designed for the scale-out of the world's data centers.”
On Tuesday, Nvidia unveiled the GH200 chip - their newest addition to the AI hardware space - in a bid to maintain their leading market position versus competitors such as AMD, Google and Amazon. Estimates place their market share in the AI chip space at around 80%. Nvidia is well-known for specializing in Graphics Processing Units (GPUs), which have proven to be the best choice for powering larger AI models such as Google's BERT and OpenAI's ChatGPT. Despite their standing, Nvidia's chips are in high demand as companies, cloud providers, and startups compete for GPU capacity to build their own AI models.
The GH200 chip is equipped with the same GPU as Nvidia's top-of-the-line AI chip, the H100. It also boasts 141GB of cutting-edge memory and a 72-core ARM central processor. During the presentation, Nvidia CEO Jensen Huang declared: "We're giving this processor a boost". He continued, "This processor is designed for the scale-out of the world's data centers". The chip will be available for purchase in the second quarter of 2021, and can be sampled by the end of the year.
AI models are typically split into two parts: training and inference. Training requires the utilization of large amounts of data and can take months to complete. This process is often done using a cluster of thousands of GPUs, such as Nvidia's H100 and A100 chips. After a model is trained, it can be used to make predictions or generate content (inference). Inference still requires a large amount of processing power, but it is used more frequently than training. With the addition of the GH200, Nvidia has enabled AI models to stay in resident on a single GPU, alleviating the need for multiple systems or multiple GPUs. As a result, the inference cost of large language models is expected to drop significantly.
Nvidia's new chip announcement comes shortly after AMD announced their own AI-oriented chip, the MI300X. The MI300X offers 192GB of memory and is marketed for AI inference. Other tech giants such as Google and Amazon have also created custom AI chips specifically for inference.
Comments