CS-4 doubles speed using the exact same chip as CS-3
cerebras launched the CS-4 system, doubling inference performance to 4,400 tokens/sec on the same 5nm WSE-3 chip by increasing power and cooling. this challenges Nvidia's dominance in high-speed inference.
cerebras unveiled the CS-4 system, delivering 4,400 tokens per second by aggressively overclocking its existing 5nm WSE-3 chip with a new 'backpack' modular design that allows three wafers per rack. while the memory remains at 44 GB, the architecture eliminates external networking bottlenecks to offer speeds claimed to be 30 times faster than Nvidia GPU setups. the system supports disaggregated inference with partners like AMD and AWS Trainium, signaling a shift toward optimizing the entire system stack rather than just chip density. this update could pressure traditional gpu manufacturers and accelerate adoption of wafer-scale engines for large-scale inference.