Cerebras has introduced the CS-4, its first multi-wafer system integrating three large wafer-scale processors to accelerate AI inference workloads. Launching this quarter, the CS-4 aims to provide up to 30 times faster processing than traditional GPU systems while operating with significantly lower power consumption.
- CS-4 integrates three wafer-scale processors for 750 petaflops compute
- System doubles clock speed of prior WSE-3 chip, cutting latency drastically
- Power efficiency reportedly halves energy use compared to comparable GPU racks
What happened
Cerebras launched the CS-4, the first system to house three of its large wafer-scale processors in a single rack chassis. Announced at the company's Supernova event, the CS-4 is designed for advanced AI inference tasks and is slated to begin shipping within the current quarter. The key chip inside, known as the WSE-3 Turbo, is based on the same core silicon as its predecessor but runs at twice the clock speed, doubling compute performance per wafer and significantly reducing latency from five to two microseconds between wafers.
This architecture delivers 750 petaflops of sparse FP16 compute power and memory bandwidth exceeding 129 petabytes per second, enabling support for AI models with over 50 trillion parameters. Additionally, Cerebras has innovated the hardware design by relocating power conversion much closer to the processors, resulting in a removable power pack at the rear and enhancing overall system efficiency.
Why it matters
The CS-4 represents a substantial leap in AI inference capabilities, promising up to 30 times faster processing on certain large models compared to traditional GPU-based systems. This speed advantage could translate to higher productivity in AI workloads and enable more complex agentic AI functions that require extended reasoning, verification, or multi-tool engagement. These improvements come at a time when speed, latency, and energy efficiency are critical factors for AI infrastructure providers and enterprises deploying frontier AI models.
Moreover, the system’s power consumption is estimated to be about half that of comparable GPU racks from competitors like AMD and Nvidia, which may offer a compelling operating cost advantage. Cerebras positions this power efficiency as a foundation for achieving tenfold increases in throughput per watt, supporting the growing demand for greener and more cost-effective AI hardware solutions.
What to watch next
The immediate focus is on the adoption of the CS-4 as shipments begin later this quarter and its performance in real-world AI inference environments, specifically with major customers such as OpenAI, AWS, and key AI institutions. Cerebras’ ability to secure customer contracts and scale deployment will be critical to validate the system’s unique multi-wafer design and speed claims.
Looking ahead, Cerebras is expected to unveil a fully new generation of silicon in 2027 that will replace the current clock-speed-enhanced processors with a redesigned chip architecture. This next generation will be closely watched to see how it advances beyond the CS-4’s performance gains and whether it can address ongoing margin pressures and revenue concentration risks posed by the company’s reliance on a small number of large customers.