The competition in the AI chip market is heating up, as Cerebras, a leading AI chipmaker, has unveiled its latest rack-scale system, the Cerebras CS-4. This innovative system boasts 30 times the tokens per second per user compared to traditional graphics processing units (GPUs).

The Cerebras CS-4: A Game-Changer in AI Acceleration

The Cerebras CS-4 is a powerful system that packs three wafer-scale WSE-3 Turbo processors, which are the largest AI semiconductors ever built. These processors boast an impressive 4 trillion transistors, making them a powerhouse in AI acceleration.

Cerebras’ systems are specifically designed for AI inference, running AI models rather than training them. Unlike traditional systems that use dynamic random-access memory (DRAM), the Cerebras CS-4 utilizes static random-access memory (SRAM). SRAM is significantly faster than DRAM but is also more complex, making it larger and more expensive. However, Cerebras’ dinner plate-sized wafers provide the necessary space to leverage SRAM, making it possible to harness its full potential.

Another key advantage of the Cerebras CS-4 is its design, which allows data to travel shorter distances compared to traditional systems that pair multiple chips. This design feature enables the Cerebras CS-4 to deliver industry-leading speeds on the largest frontier models, fundamentally changing the paradigm in AI acceleration.