Cerebras says latest offering is ‘fastest AI accelerator in the industry’ as it takes aim at Nvidia

The competition is heating up for Nvidia (NVDA), as AI chipmaker Cerebras (CBRS) unveiled its newest rack-scale system, which it claims offers 30x the tokens per second per user compared with graphics processing units.

Called the Cerebras CS-4, the system packs three wafer-scale WSE-3 Turbo processors, which the company says are the largest AI semiconductors ever built, packing 4 trillion transistors.

Cerebras’ systems are designed for AI inference, running AI models, rather than training them. They use a kind of memory called static random-access memory (SRAM) rather than dynamic random-access memory (DRAM).

SRAM is far faster than DRAM, but much more complex, making it larger and more expensive than DRAM. That makes using SRAM all but impossible with your average pint-sized processor.

Cerebras’ dinner plate-sized wafers provide the space needed to leverage SRAM.

What’s more, the fact that Cerebras’ processors are a single product means data has to travel shorter distances than, say, Nvidia or AMD’s systems, which pair multiple chips that have to move data around more.

“Historically, fast inference meant using smaller and less capable models. Cerebras CS-4 delivers industry-leading speeds on the largest frontier models, fundamentally changing the paradigm,” Cerebras CEO Andrew Feldman said in a statement.

“Every aspect of the design…