Nvidia challenger Cerebras teams with Dell on generative AI

It remains to be seen if there is such a thing as a realistic contender to Nvidia’s AI throne, but Cerebras Systems is shooting its shot. The company has been claiming that its CS-3 system, based on its Wafer Scale Engine 3 (WSE-3) processor launched in March, bests GPU memory and performance, and this week Cerebras announced a collaboration with Dell Technologies that could expand its reach, and put it in contention for more enterprise generative AI projects.

Cerebras and Dell together will offer AI systems and supercomputers, white-glove large language model training, and machine learning expert services. The partners additionally said in a statement that the collaboration also includes a new memory storage solution powered by Dell PowerEdge R6615 servers and AMD EPYC 9354P CPUs for Cerebras AI supercomputers, “enabling enterprises to train models orders of magnitude larger than the current state of the art… extending [Cerebras’] 82TB streaming memory supercomputers clusters almost infinitely to train models of any size.”

Cerebras co-founder and CEO Andrew Feldman, who in the past has boasted that his company’s AI acceleration technology enables 880 times the memory capacity of GPUs, 97% less code to build large language models, push-button model scaling and better data preprocessing, said the Dell collaboration is a major turning point in the ability of the Sunnyvale, California, firm to reach a much broader market.

“This opens up our global sales distribution channels in a meaningful way, while providing customers with the additional AI hardware, software and expertise needed to enable full-scale enterprise deployments,” he said.

This week’s announcement also follows another from last month in which Cerebras said it had teamed up with researchers from Sandia, Lawrence Livermore, and Los Alamos National Laboratories to achieve a “breakthrough” in molecular dynamics simulations. Using the second-generation Cerebras Wafer Scale Engine (WSE-2) researchers performed atomic-scale simulations “at the millisecond scale – 179x faster than what is possible on the world’s leading supercomputer ‘Frontier,’ which is built with 39,000 GPUs,” according to a Cerebras statement.

While many companies have been positioning to get a piece of an AI chip market that Nvidia GPUs have been dominating early on, there could be something to Cerebras’ claims about its wafer-scale technology that sets it apart, according to Jack Gold, president and principal analyst at J. Gold Associates.

“Cerebras is doing some state-of-the-art processing for AI,” Gold told Fierce Electronics. “Its wafer-scale processing systems mean that you can have a lot more compute running the AI without having to have multiple unique modules [like Nvidia H100s]. It eliminates the need for massive high-speed interconnect, and vastly increases memory utilization as it’s local to each processing node and not having to go over a bus off the chip.”

Indeed, Nvidia’s NVLink GPU interconnect technology has been a key part of its success in enabling the networking of many GPUs, and drove AMD, Intel and other firms to recently propose an alternative open standard called UALink. Meanwhile, Cerebras has touted what it can do on one large wafer.

Gold said, “Wafer-scale has a lot of advantages over the massive module model with GPUs. It uses less power, runs faster, and can run much larger parameter models. So ultimately it can scale to much larger models that GPUs will have trouble scaling out to.”

But Gold also cautioned that “it’s also a lot harder to pull off. Cerebras claims they have done so, but don’t really talk about the cost or the volume of systems they can supply. It is very hard to build a wafer-scale compute system. They claim they are well positioned on TSMC 5nm, but we’re already moved beyond 5nm. At the end of the day can they produce in large volume? That’s an important question.”

Gold acknowledged that Cerebras is doing “something special” that represents a competitive threat to “the Nvidia huge-number-of-GPU-modules solution,” but also said Cerebras “doesn’t have the visibility that Nvidia has, and it has a lot of marketing to do to catch up and surpass Nvidia.”

For now, the more admirable and realistic aim is just to provide the market with more choices, especially as Nvidia reportedly has drawn antitrust concerns of its early market dominance. Cerebras is delivering another option.

As Feldman told Fierce Electronics, “Nobody likes to be dependent on one vendor. So right now, there’s a groundswell of interest amongst our top customers to ensure there’s a healthy ecosystem with viable competition and different approaches to solving problems. There are tremendous areas of opportunity for improvement, and billions of dollars of business that need to be earned, which is both possible and important. I don’t know of any healthy markets where there’s just one top player.”

He added that Cerebras, a nine-year-old company, is just getting started. “This year we will deploy more many 10s of exaFLOPs of compute, we have won key customers around the world, and we have announced partnerships that can propel us forward, including Mayo Clinic, Qualcomm, Dell, Neural Magic, G42 and others,” Feldman said. “There still remains many opportunities across big OEMs, CSPs, international players and in many places others aren’t thinking about. At Cerebras we always want to be doing something that no one else seems to be doing. We will continue building products that are bigger and better, and we will continue partnering with the top players in the ecosystems to deploy them widely.”