Cerebras CS-3 (WSE-3): The Complete Guide

Cerebras’s own head-to-head math against NVIDIA’s then-new B200: “a single CS-3 equates to about 3.5 DGX B200 servers” on raw compute (125 vs. 36 petaflops), a 2.2x improvement in performance per watt 3 — see the performance section below for why that is not the same as CS-3 drawing less power in absolute terms. This is Cerebras’s own comparison, not an independent benchmark.

At its inference launch, Cerebras served Meta’s Llama 3.1 8B model at “1,800 tokens a second” and Llama 3.1 70B at “450 tokens a second” — speeds it attributed directly to keeping the model’s weights on the wafer’s own memory rather than fetching them from separate memory chips 2 5.

For models too large for one wafer’s on-chip memory, Cerebras pairs the chip with an external unit called MemoryX, offered in sizes from 24 TB and 36 TB (enterprise) up to 120 TB and 1,200 TB (hyperscale); the largest configuration can hold “models with 24 trillion parameters” 6. Cerebras says a full 2,048-system cluster, using its Weight Streaming architecture, “looks and programs like a single chip” 6.

Cerebras’s own comparison against a full NVIDIA NVL72 rack (72 GPUs): a single CS-3 “provides more than 200x the interconnect bandwidth” — stated as 27 petabytes per second of aggregate on-wafer fabric bandwidth across the chip’s 900,000 cores — and “over 80x the memory capacity” when paired with its largest MemoryX configuration 3. This 27 PB/s core-to-core fabric figure is a different measurement from the 21 PB/s figure above, which Cerebras states separately as memory-to-core bandwidth — both describe traffic within one wafer, not between separate chips, and Cerebras’s own materials do not explain how the two relate to each other.

Cerebras’s January 2025 blog post on this generation’s manufacturing says each wafer builds 970,000 physical cores and activates 900,000 of them, routing around the rest, and describes the chip as “about 100x more fault tolerant” than a GPU because a single core failure costs a much smaller share of the total wafer 7.

By March 2025 Cerebras said its own data centres in Santa Clara, Stockton and Dallas were already online, running on this generation, with more sites in the US and Europe planned 9. Under the AWS collaboration, Cerebras’s CS-3 handles one stage of an AI request and AWS’s own Trainium chips handle another (see the Amazon Trainium guide on this site) 8.

At its August 2024 inference launch, Cerebras priced Llama 3.1 8B at 10 cents and Llama 3.1 70B at 60 cents per million tokens — the earliest published Cerebras Inference prices, all served from CS-3 hardware 2.

This is Cerebras’s own published comparison, not an independent third-party benchmark, and its “half the power consumption” framing needs a caveat. By the same source’s own figures, CS-3 draws up to 23 kW peak against a DGX B200’s 14.3 kW — more absolute power, not less. “Half” refers to power per unit of compute (performance per watt), which is where the 2.2x figure comes from; it is not a claim that a CS-3 system pulls fewer watts off the wall 3. Cerebras also claims a full 2,048-system CS-3 cluster can “train Llama2-70B from scratch in less than a day” versus “approximately a month” it attributes to Meta’s own GPU infrastructure for the same task 6 — a claim about a specific competitor’s reported training time, not a controlled side-by-side test Cerebras ran itself.

WSE-3 launched almost three years after WSE-2 (April 2021 to March 2024), a longer gap than the roughly 20 months between the first two generations 10 1. Its successor, CS-4, was announced in August 2026 with Cerebras saying “first CS-4 shipments begin this quarter” — language this site treats as a projection, not a confirmed shipment, until an actual customer delivery is reported.

The August 2024 inference launch reframed this chip’s purpose in Cerebras’s own business: WSE-3 was announced in March 2024 as a training chip, competing on raw petaflops against NVIDIA’s training GPUs 1 3, but within five months Cerebras was marketing the same silicon’s speed advantage for serving finished models to users, not just training them 2 — the same hardware, sold against two different competitive claims within one calendar year.