On Monday, Nvidia launched the H200, a GPU created for instructing and utilizing the kinds of AI models fueling the generative AI explosion. The H200 features 141GB of recent "HBM3" memory which will aid it in constructing texts, illustrations or forecasts via AI models. The enthusiasm for Nvidia's AI GPUs has been a major boost for the firm, with revenue predicted to soar by 170% this quarter.
On Monday, Nvidia unveiled the H200, a graphics processing unit designed to train and deploy artificial intelligence models that are driving the generative AI surge. The H100, that OpenAI employed to train its most sophisticated large language model, GPT-4, is an upgrade from the chip. Big companies, startups, and government agencies are all competing for limited availability of the chips. According to an estimate from Raymond James, the H100 chips cost between $25,000 and $40,000, and thousands of them need to be combined to create the most significant models in a process termed "training." The hype surrounding Nvidia's AI GPUs has helped push the company's stock to a year-to-date increase of more than 230%. Nvidia anticipates nearly $16 billion in revenue in its fiscal third quarter, which is a 170% increase from the previous year. The key advancement with the H200 is that it features 141GB of next-generation "HBM3" memory, which can help the chip perform "inference," or employ a large model after training to produce text, images, or predictions. Per Nvidia, the H200 will generate output nearly twice as quickly as the H100, based on a test utilizing Meta's Llama 2 LLM. The H200 is set to ship in the second quarter of 2024 and will rival AMD's MI300X GPU. AMD's chip, similar to the H200, has extra memory over its predecessors, which can help fit enormous models on the hardware to conduct inference.
Nvidia has stated that the H200 is compatible with the H100, which means AI businesses that are already running with the previous model will not need to modify their server systems or software to use the newer version. This is available in four-GPU and eight-GPU server configurations on the company's HGX full systems, as well as in an integrated chip called GH200, which couples the H200 GPU with an Arm-based processor. However, the H200's reign of being the fastest AI chip may be short-lived. While many configurations of chips are provided by Nvidia, every two years, semiconductor makers create new architectures that can bring in much more performance compared to just adding memory or optimizing. Both the H100 and H200 are based on Nvidia's Hopper architecture. In October, Nvidia told its investors they would switch to a one-year release cadence, due to the large demand for its GPUs. The company displayed a slide hinting that it will announce and launch the B100 chip, based on the upcoming Blackwell architecture, in 2024.WATCH: We're a big believer in the AI trend going into next year.
Don't miss our stories from CNBC PRO: Find out which five stocks make up 75% of Warren Buffett's equity portfolioTwo financial institutions just increased their 1-year CD yield to 5.3%Learn a smart strategy for capitalizing on Apple's post-earnings dipStifel forecasts that the S&P 500 will keep on climbing and reach 4,400 - it suggests 10 stocks to take advantage of the upswing.
Comments