Cerebras Systems has unveiled the CS-4, which it describes as the fastest artificial intelligence (AI) accelerator in the industry.
The system was launched at the company's Supernova event on Tuesday and will begin shipping this quarter.
Cerebras, the Silicon Valley chipmaker that floated on the Nasdaq in May, builds unusually large processors known as wafer-scale engines rather than clusters of smaller graphics chips.
The CS-4 packs three of these processors into a single rack and is aimed at inference, the stage at which a trained model generates responses to user queries.
The company claims the system runs frontier models up to 30 times faster than rival systems based on graphics processing units, the chips that dominate AI computing.
It also cites up to twice the performance of the previous CS-3 and up to 10 times more throughput per watt.
Cerebras puts output at more than 4,400 tokens, the units of text a model generates, per second for each user on one open-source model.
Same silicon, more power
The headline gains do not come from a new chip.
Each CS-4 uses the same processor as the CS-3, built on Taiwan Semiconductor Manufacturing Company's 5-nanometre process.
The main change is a clock speed roughly doubled to 2.8 gigahertz.
Independent analysts noted that per-wafer compute and memory bandwidth both exactly double, the signature of a clock increase rather than a redesign.
Cerebras has a genuinely new generation, the CS-5, scheduled for 2027.
Where the design does look distinctive is efficiency, with a rack that draws roughly half the power of comparable systems from Nvidia and AMD.
The company says the new architecture uses 50% fewer components, cuts manufacturing time by 60%, and reduces deployment from days to hours.
Testing demand
The launch is the first test of demand since Cerebras raised $5.55 billion in the largest US technology listing since Snowflake.
Faster, cheaper-to-build systems could lift both revenue and margin per unit while allowing the company to fill orders more quickly.
Cerebras already counts OpenAI and Amazon among its customers, and the CS-4 is designed to run OpenAI's new Ultrafast mode, which serves its most capable models at higher speed.
The commercial question is whether wafer-scale economics can hold up as Nvidia and AMD push their own rack-scale designs.