Nvidia has strong competitors in the field of AI accelerators, but it’s too early to write it off

Amazon, AMD and a number of startups are increasingly offering alternatives to Nvidia artificial intelligence accelerators, but it is too early to write off the market leader – it still offers undeniable advantages in performance and software, writes The New York Times.

Image source: nvidia.com

A year ago, AMD introduced the Instinct MI300 AI accelerators, which, according to preliminary estimates, have now reached $5 billion in sales. Amazon this week introduced the second generation of its Trainium accelerators, which will help the company strengthen its position in the field of training and running AI systems. Judging by customer reactions to these solutions, Nvidia has worthy alternatives. The company has long dominated the market, helping it grow its value to $3 trillion, but now its competitors are showing they can deliver faster speeds at lower prices.

This is because a number of technology companies, including not only giants like AMD and Amazon, but also small startups, have begun adapting chips to run already trained AI (inference) models. “True commercial value comes from inference, and inference is starting to gain scale. We are starting to see changes,” said Qualcomm CEO Cristiano Amon; The mobile chip maker is also planning to use Amazon accelerators for AI tasks. Nvidia’s competitors decided to follow its example in another way – they began to offer not just accelerators, but ready-made computers designed to work with AI.

Increasing competition in the field of AI hardware became obvious when Amazon offered customers rental resources based on the new Trainium2 accelerators – including Apple, which left positive reviews. Amazon also introduced 16- and 64-accelerator servers with high-speed connections that help boost performance when running AI systems. Amazon is also building a giant AI cluster for its partner Anthropic, which will be able to train its models on hundreds of thousands of Trainium2 accelerators. Spending by data center operators on AI systems without Nvidia chips will grow by 49% this year to reach $126 billion, Omdia analysts estimate.

Image source: aws.amazon.com

But increased competition does not yet mean that Nvidia will lose its leadership. The company’s CEO, Jensen Huang, said it has strengths in AI software and model-running capabilities. The new Blackwell family of chips require more power to operate, but their performance per watt is higher than that of the competition. Current Nvidia accelerators cost up to $15,000 apiece, and prices for Blackwell are expected to be in the tens of thousands. Meanwhile, startups that have managed to attract investment are rising, with SambaNova Systems, Groq and Cerebras Systems claiming to have achieved both performance and cost advantages in their AI launches.

Some clients have already begun to change their work patterns. Texas Advanced Computing Center executive director Dan Stanzione said the organization plans to buy a Blackwell-based supercomputer next year, but will use SambaNova accelerators, which have lower power consumption, to run AI tasks. energy and price. Meta✴ also reported that it trained the Llama 3.1 405B model on Nvidia chips, but uses AMD MI300s accelerators to run it.

Amazon, Google, Microsoft and Meta✴ continue to build large clusters on Nvidia accelerators, but are also developing their own chips to speed up certain computing tasks to reduce costs. Google intends to begin leasing resources on its sixth-generation Trillium accelerators, which are almost five times faster than their predecessors, by the end of the year. Amazon, considered an AI laggard, has committed $75 billion to AI accelerators and other computing hardware this year. The first-generation Trainium accelerators have not received much market acceptance, but the company is more optimistic about Trainium2, with more expected to be released within a year. more powerful Trainium3. Experts estimate that Trainium2 will offer a 40% increase in performance per dollar compared to Nvidia-based hardware.