In recent years, Nvidia has been a major beneficiary of the generative Artificial Intelligence (AI) boom. Companies like OpenAI have built their most advanced AI models using clusters of Nvidia Graphical Processing Units (GPUs), and the firm’s latest AI chips are highly sought after as the sector continues to grow.
The firm has responded by doubling down on the AI market, accelerating the development of faster, more powerful GPUs. But with rivals increasingly eying Nvidia’s crown, can the chipmaker hold onto its lead in the face of rising competition?
Building on the success of its predecessor, the H100, on Monday, November 13, Nvidia unveiled its latest GPU for the AI market: the H200.
“With Nvidia H200, the industry’s leading end-to-end AI supercomputing platform just got faster,” boasted the company’s accelerated computing lead Ian Buck.
Compared to H100 chips, the new GPUs will offer nearly double the inference speed, a crucial metric that determines how fast machine learning models can process data.
Once the new chips ship next year, they will be installed in data centers operated by Google, Microsoft and Amazon Web Services (AWS), as well as the specialist AI cloud providers CoreWeave, Lambda and Vultr.
But Nvidia isn’t resting on its laurels. Before the first H200 has even rolled off the production line, the company teased the next generation of AI chips at the recent Super Computing 2023 conference.
Before the end of 2024, the firm will launch the B100, delivering yet another performance boost to its range of AI-focused GPUs.
According to a presentation slide from the event, using OpenAI’s GPT-3 large language model for comparison, the H200 will perform 18x better than the current H100s, while the B100 range is set to deliver yet another leap in processing capacity.
As things stand, Nvidia’s H100 is the most powerful GPU on the market and the go-to AI chip for cloud data centers and purpose-built supercomputers. But with rival chipmakers on the move, it’s no wonder the firm has accelerated its new product roadmap.
With Nvidia’s current monopoly of the data center GPU market in its sights, the South Korean semiconductor manufacturer Sapeon unveiled its latest AI processor during SK Group’s annual tech summit last week.
According to the company’s CTO, Michael Shebanow, the new X330 chip is 2 times faster and 1.3 times more energy efficient than a “key rival” that is likely a reference to Nvidia’s H100.
“We‘re focused on being very good at AI inference,” Shebanow remarked , making it clear that Sapeon is vying for a share of the expanding AI chip market.
What’s more, just as Sapeon is preparing to challenge Nvidia’s GPU dominance, Intel is doubling down on its AI offering, hoping to salvage a role for CPUs in the booming AI data center business.
Commenting on Intel’s latest Xeon processors during a recent Q3 earnings call, CEO Pat Gelsinger said the new chips position the firm to win back lost share in the data center market.
“We expect to capture a growing portion of the accelerator market in 2024 with our suite of AI accelerators,” he noted.
While GPUs have been instrumental in training the largest contemporary AI models, as the sector grows, the day-to-day operation of existing models will become more important. Here, cheaper, more energy-efficient CPUs could make their comeback.
For example, although OpenAI’s GPT models were trained on a custom-built cluster of Nvidia GPUs, when it comes to making GPT-based services available to Azure customers, Microsoft is equipping its data centers with new “Maia” CPUs launched on Wednesday, November 15.
As well as being cheaper than purchasing GPUs from Nvidia, the new CPUs are also cheaper to run, delivering more computing power per unit of energy consumed.
Commenting on the development, Sam Altman, who until recently was OpenAI’s CEO, said the new chips will contribute toward “making [AI] models cheaper for our customers.”
Aside from GPUs and CPUs, there is a third type of processor that promises to transform AI development.
First developed by Google, Tensor Processing Units (TPUs) are currently best suited for tasks such as image processing that require a high volume of low-precision computations. However, the firm continues to develop the technology, which it envisages as an alternative to GPUs for a growing number of machine learning workloads.
Having now left OpenAI, one of Altman’s prospective new ventures is Tigris – a TPU startup he has reportedly been discussing with investors. According to people familiar with the matter cited by the New York Times, the new company would compete with Nvidia in the AI chip market.