Cerebras unveils WSE-3 Turbo with ten times greater capacity

Artificial intelligence accelerator manufacturer Cerebras announced the launch of the WSE-3 Turbo chip and the CS-4 system. The new system delivers ten times greater processing capacity compared to the earlier CS-3 model while doubling performance speed. The WSE-3 Turbo chip maintains 900 thousand cores and 44 gigabytes of cache memory but doubles performance in reduced-precision floating-point computation, memory bandwidth, inter-chip interconnection, and input/output bandwidth. The CS-4 can integrate up to three WSE-3 Turbo chips to increase computational density. In tests with open-source language models, the system delivered 4,465 tokens per second, being 30 times faster than GPU-based solutions and 93% faster than its predecessor. For models with extreme complexity, the ultra-low latency of 2 microseconds enables processing over 1,000 tokens per second at a scale of 10 trillion parameters. The system supports native distribution of inference workloads, functioning as a decoding unit integrated with heterogeneous hardware. Built on the Nexus platform, the CS-4 reduces the component count by 50% and virtually eliminates power losses at the board level, significantly improving energy efficiency.