Cerebras releases CS-4 accelerator: 4 trillion transistors + 900,000 cores, 30 times faster than Nvidia GPU. AI chip company Cerebras Systems officially launched a new generation of rack-level AI system CS-4 on August 18, local time, claiming that it is the "fastest AI accelerator in the industry" with the highest inference speed.

Cerebras releases CS-4 accelerator: 4 trillion transistors + 900,000 cores, 30 times faster than Nvidia GPU. AI chip company Cerebras Systems officially launched a new generation of rack-level AI system CS-4 on August 18, local time, claiming that it is "the fastest AI accelerator in the industry" and can infer inference speeds up to 30 times that of GPU solutions. The accelerator is driven by three new WSE-3 Turbo wafer-level processors, each integrating 4 trillion transistors and 900,000 AI-optimized cores, covering 46,225 square millimeters of silicon area. CS-4 is the first product of Cerebras' new generation Nexus rack-scale platform architecture. Using a modular design to separate computing, power supply and I/O into independent components, the number of parts is reduced by 50% compared with the previous generation. The power conversion module of this product is shortened from about 50 mm away from the processor on the traditional GPU motherboard to about 0.5 mm. The pluggable backpack design integrates power conversion, liquid cooling and control electronics, reducing deployment time from days to hours. In terms of performance, CS-4 provides 750 PFLOPs of AI computing power, 129.6 PB of memory bandwidth per second, and 7.2 Tbps of I/O throughput. Compared with the previous generation CS-3, the speed is doubled and the throughput per watt is increased by 10 times. In the GPT-OSS-120B model test, CS-4 can generate more than 4,400 tokens per user per second, up to 30 times that of the GPU solution, and the inter-wafer delay is reduced from about 5 microseconds to 2 microseconds. CS-4 is built using TSMC’s 5nm process and system-on-wafer (SoW) packaging. Its biggest feature is that it does not require high-bandwidth memory and is replaced by SRAM. A single cabinet can accommodate 3 wafers. The system supports models with more than 50 trillion parameters and a TDP of about 125 to 135kW. Shipments are expected to begin in the third quarter of 2026. Cerebras works with both Amazon AWS and AMD to allow customers to outsource the highly compute-intensive pre-population stage to AMD Instinct GPUs. via cnBeta.COM - Chinese industry information station (author: Source: Kuai Technology)