Cerebras touts huge inference gains
Cerebras claims its CS‑3 delivers up to 20x faster AI inference than GPUs, with 2,500+ tokens/sec and 44GB on‑chip SRAM that eliminates HBM needs reported. Those specs feed a growing wafer‑scale vs. chiplet debate for inference‑heavy deployments.
AWS and Cerebras announced a partnership on March 13, 2026 to deploy a disaggregated inference stack that pairs AWS Trainium for prefill with Cerebras CS‑3 for decode on Amazon Bedrock, connected via Elastic Fabric Adapter. (press.aboutamazon.com) Cerebras’ WSE‑3 wafer-scale engine is described in its filings as a 5nm design with roughly 4 trillion transistors, about 900,000 AI‑optimized cores and a peak of 125 petaflops per wafer, with system clusters scaling to 2,048 CS‑3 nodes. (cerebras.ai) Multiple technical summaries and independent analyses report aggregate on‑chip memory bandwidth in the WSE‑3 at roughly 21 petabytes/sec, a figure used to justify eliminating external HBM transfers for large models. (arxiv.org) Hardware reviewers have equated a single WSE‑3’s raw compute to dozens of H100‑class GPUs—Tom’s Hardware estimated an equivalence near ~62 H100s for certain workloads—while Cerebras cites single‑chip cluster behavior for very large models. (tomshardware.com) Cerebras contrasts its wafer‑scale approach with streaming and chiplet strategies from rivals (Groq’s LPU, SambaNova’s RDFUs and chiplet‑based GPU stacks), arguing that on‑wafer fabric and massive on‑chip bandwidth favor high‑throughput decode stages even as competitors emphasize latency and modular capacity. (cerebras.ai) Early commercial rollouts reflect that strategy: Cerebras says CS‑3s are shipping to customers and powering the Condor Galaxy 3 deployment with partner G42 (operational Q2 2024), and AWS positions the Trainium+CS‑3 Bedrock offering as live in data centers in the coming months. (cerebras.ai)
