Cerebras has launched the CS-4, a rack-scale AI accelerator that claims the title of the fastest system in the industry. This unit sits in a single server cabinet containing compute units, power supplies, and cooling systems. Although it continues to use the 5nm WSE-3 chip from the previous generation, the new hardware doubles performance by increasing clock speeds through enhanced power delivery and improved thermal management. A single rack now accommodates three wafers rather than two, delivering up to 4,400 tokens per second per user. The company states this output is up to 30 times faster than configurations running on Nvidia GPUs. Memory capacity remains unchanged at 44 GB per wafer.
The significance lies in the ability to achieve higher throughput without changing the underlying silicon process. Cerebras employs a modular Backpack design to speed up assembly and supports disaggregated inference through partners such as AMD and AWS Trainium. Analysts at SemiAnalysis note that networking gains appear modest, with further details expected at the Hot Chips conference. OpenAI uses Cerebras hardware for Codex Spark among other projects.
* Three wafers fit per rack compared to two previously
* Clock speeds increase via better cooling and power
* Memory capacity stays fixed at 44 GB per wafer
