Cerebras Unveils CS-4: Chip Stays Put, Inference Claimed to Be 30x Faster Than GPU Solution

Perceive Beating AI News: AI chip company Cerebras has released the new generation AI inference server CS-4. The name has been upgraded from CS-3 to CS-4, but the core still uses the WSE-3 series chip, without introducing a new WSE-4. The main upgrades come from the overall system design, with each CS-4 integrating 3 WSE-3 Turbo chips and undergoing optimizations in power delivery, cooling, and chip-to-chip communication.

Cerebras stated that the inference speed of CS-4 can reach up to 30 times that of a traditional GPU solution, with the highest unit power throughput being 10 times that of CS-3. The company did not specify which GPU the "30 times" is specifically compared to, and the performance gap varies significantly across different models and configurations, so this number is best viewed as the company's claimed best-case scenario.

CS-4 has also doubled the single-chip inference speed compared to the previous generation. Cerebras mentioned that even for models with 1 trillion parameters, it can achieve over 1000 tokens/second. However, this result is extrapolated from internal testing and not from actual benchmarking of a 1 trillion parameter model.

The first batch of CS-4 servers is set to be delivered in the third quarter of this year, and Cerebras plans to release the next generation of chips and servers in 2027.