Cerebras Launches CS-4: Up to 30x Faster Inference Than GPUs
Original: Cerebras CS-4
Why This Matters
CS-4 targets hyperscale inference for trillion-parameter models, a critical bottleneck in frontier AI deployment.
Cerebras has announced the CS-4, a rack-scale AI accelerator featuring three WSE-3 Turbo wafers per system. It delivers up to 30x faster inference than GPU systems and 10x more throughput per watt than the previous CS-3. First shipments begin this quarter.
Cerebras has unveiled the CS-4, its next-generation AI accelerator built around the new Nexus Platform Architecture. The system houses three WSE-3 Turbo wafers, with each wafer delivering up to 2x the speed of the previous generation's WSE-3. Key performance claims include up to 30x faster inference compared to production GPU systems, over 1,000 tokens per second on models exceeding 10 trillion parameters, and wafer-to-wafer interconnect latency as low as 2 microseconds.
A central design innovation is the 'Wafer-Scale Backpack,' a self-contained modular unit that integrates the wafer, power conversion, direct liquid cooling, high-speed I/O, and control electronics into a compact 3D package using 50% fewer components. Power delivery is placed just 0.5mm from the processor—approximately 100x closer than conventional GPU boards—nearly eliminating board-level power loss and enabling twice the power delivery to the WSE-3T.
The CS-4's Nexus architecture separates the stable power, cooling, and networking layer (the PowerRack) from the compute backpacks, allowing infrastructure to be installed and facility-qualified before compute units arrive. Cerebras states this reduces deployment time from days to hours. First CS-4 shipments are scheduled to begin in the current quarter.