Cerebras's Next Generation CS-4: Fast Just Got Faster

Key Points:

• Cerebras unveiled the CS-4, its fourth-generation rack using the same 5nm WSE-3 wafer-scale engine as CS-3 but doubling performance by increasing clock speed and power consumption, while maintaining similar cost and memory capacity per wafer.
• The CS-4 features a modular "backpack" rack design that separates power delivery and compute, allowing three wafers per rack (up from two in CS-3), improved manufacturability, easier deployment, and a total power draw near double that of CS-3.
• Networking enhancements include doubled off-wafer I/O bandwidth to 2.4Tb/s, a new field-upgradeable I/O module enabling open, heterogeneous disaggregated inference architectures, and reduced network latency down to 2-3 microseconds with direct wafer-to-wafer links.
• Cerebras emphasizes pipeline parallelism for inference due to limited on-wafer SRAM (44GB), focusing on high interactivity/low throughput workloads, while planning disaggregated setups with partners like AMD and AWS to overcome memory constraints by pairing CS-4 with HBM-based systems.
• The company is developing the next-generation wafer-scale engine and rack platform "Nexus," targeting 2x performance improvement annually and 20x throughput by 2027, alongside investments in reliability, error recovery, and yield improvements.