Cerebras has introduced the CS-4 AI accelerator, which CEO Andrew Feldman calls the fastest system in the industry. The CS-4 is a rack-scale product: a full server cabinet that goes into a data center as a single unit, complete with compute units, power and cooling.

What is new and what is not

The interesting part of this announcement is where the performance gain does not come from. The CS-4 still runs the 5nm WSE-3 chip. It doubles the performance of the earlier CS-3 by boosting clock speed through more power and better cooling.

What changed:

  • Wafers per rack — three now, instead of two.
  • Per-user throughput — up to 4,400 tokens per second.
  • Modular design — a new layout called "Backpack" for faster assembly.
  • Disaggregated inference — through partners including AMD and AWS Trainium.

What did not change is memory: it stays at 44 GB per wafer.

Background: what wafer-scale means

What separates Cerebras from the rest of the industry is that it does not cut the silicon wafer. In normal manufacturing a wafer is diced into dozens of separate chips, which are then wired together on a board. Cerebras leaves the wafer whole and uses it as one enormous processor.

The payoff is that chip-to-chip communication disappears. On large models the bottleneck is often not computation but moving data between processors. With the wafer intact, that traffic stays inside the silicon and moves far faster.

The cost is memory. On-wafer memory is fast but limited — here, 44 GB per wafer. For comparison, the weights of a large language model can occupy many times that. Which is why memory staying flat is the constraint that decides what workloads this system suits.

The comparison claim, and a caveat

Cerebras says this is up to 30 times faster than setups built on Nvidia GPUs. Ratios like that deserve careful reading: vendor comparisons are usually run on workloads where their own architecture is strongest, and the other side's configuration goes unstated.

Analysts at SemiAnalysis also see the networking gains as fairly small. More details are expected at the Hot Chips conference.

Why it matters

This announcement shows where attention is shifting in AI hardware. For years the industry talked about fabrication processes: how many nanometers, how many transistors. By doubling performance on the same chip, the CS-4 suggests the gain can now come from the system around the silicon rather than the silicon itself — power delivery, cooling capacity, rack density.

The practical consequence is that performance can rise without waiting for a new process node. But there is a cost. Raising clock speed through power and cooling raises electricity draw per rack. In a period when the binding constraint on the data center side is megawatts, that means moving the bill from one column to another.

Cerebras hardware is used by OpenAI for Codex Spark, among others. The company's focus on inference speed is another sign of weight shifting in the industry from training toward inference.