Cerebras Unveils WSE-3 Turbo Chip and CS-4 System with 10x Token Capacity
Cerebras has announced the WSE-3 Turbo chip and the CS-4 system, which delivers a tenfold increase in token capacity and doubles the compute performance and bandwidth compared to the previous generation. The CS-4 system can integrate three WSE-3 Turbo chips and achieves a token delivery speed of 4,465 tokens/second on the OpenAI GPT-OSS model. This release significantly advances wafer-scale AI hardware, offering a massive speedup over traditional GPU solutions for large language model (LLM) inference. It enables extremely low-latency wafer-to-wafer interconnects, allowing ultra-large 10-trillion-parameter models to run at high speeds. While maintaining 900,000 cores and 44GB SRAM, the WSE-3 Turbo doubles FP16 sparse AI compute, memory bandwidth, and I/O bandwidth. The CS-4 system natively supports workload splitting, allowing it to act as a decoding unit paired with heterogeneous prefill hardware.
## BACKGROUND
Wafer-scale integration (WSI) involves building very large integrated circuits from an entire silicon wafer to create a single "super-chip," which is much larger than traditional GPUs. Additionally, exploiting sparsity in AI hardware allows systems to skip zero-value computations, significantly accelerating deep learning inference and training.