Cerebras has launched the CS-4 AI accelerator, which the company claims is the fastest system in the industry. The CS-4 is a rack-scale product, designed to fit into data centers as a single unit, complete with compute units, power, and cooling. The chip continues to use the 5nm WSE-3 architecture but achieves double the performance of the earlier CS-3 by increasing clock speed through enhanced power and cooling. A single rack now holds three wafers instead of two, delivering up to 4,400 tokens per second per user, which Cerebras states is up to 30 times faster than setups using Nvidia GPUs. Memory capacity remains at 44 GB per wafer. Cerebras is also introducing a new modular 'Backpack' design for faster assembly and is collaborating with partners like AMD and AWS Trainium for disaggregated inference. Analysts at SemiAnalysis noted that the networking improvements are relatively modest, and more details are expected at the Hot Chips conference. Cerebras hardware is used by OpenAI for Codex Spark, among other applications.
Cerebras emphasized that the CS-4 maintains the same 5nm WSE-3 chip as the CS-3 but improves performance by boosting clock speed with increased power and better cooling. The company claims the new system can process up to 4,400 tokens per second per user, a significant increase from previous models. The CS-4 also retains a memory capacity of 44 GB per wafer, which remains unchanged from its predecessor. The modular 'Backpack' design is intended to streamline manufacturing and assembly processes, allowing for quicker deployment of the hardware. Cerebras is also working with partners such as AMD and AWS Trainium to enable disaggregated inference, which could improve flexibility in AI workloads. However, analysts at SemiAnalysis suggested that the networking enhancements are relatively minor, indicating that the primary improvements lie in the hardware itself.
Cerebras hardware is currently used by OpenAI for Codex Spark, according to the source. The company plans to provide more details about the CS-4 at the Hot Chips conference, where further specifications and performance benchmarks are expected to be shared. The CS-4 represents an evolution in Cerebras' rack-scale AI accelerators, offering enhanced performance without a complete redesign of the underlying chip architecture.
Source: thedecoder
