Cerebras WSE-2
AI accelerator built from an entire silicon wafer
The Cerebras WSE-2 is a massive artificial intelligence processor that uses a whole silicon wafer as a single chip. Designed to handle growing neural network models, it integrates compute and memory on a vast scale, supported by external systems for storage and networking.
- Uses an entire silicon wafer as a single chip
- Includes 40GB of onboard SRAM memory
- Supports sparse compute for efficiency gains
- Scales via MemoryX and SwarmX interconnects
- Successor to the original WSE-1 engine
What it is
The Cerebras WSE-2 is an artificial intelligence accelerator that differs from conventional chips by using an entire silicon wafer as a single integrated circuit. Most processors are cut from wafers into small individual dies, but this device keeps the whole structure intact to maximise connectivity and memory bandwidth. It resides within the CS-2 system and targets large-scale neural network training where model size often exceeds available memory. The design addresses constraints in data movement by keeping compute resources physically close to stored weights.
How it came about
Cerebras previously launched the WSE-1 and its accompanying CS-1 system, proving that a full wafer could function as one chip. The WSE-2 represents an evolution of this concept, aiming for higher performance and better scaling capabilities. Presented at Hot Chips 33, the second generation focuses on handling the rapid growth of neural network models. The company observed that memory capacity is a primary bottleneck for modern AI workloads, driving the need for larger on-chip resources and efficient external expansion.
Inside the chip
The architecture prioritises sparse compute, allowing the hardware to skip zero-value calculations to improve efficiency. It includes 40GB of onboard static random-access memory (SRAM) for fast data access. To manage models that exceed this local capacity, the system uses Weight Streaming to move data between the chip and external storage. This approach enables the processing of extremely large network layers without stalling the compute units, leveraging sparsity to gain performance improvements similar to those sought by other industry players.
How it was made
The manufacturing process creates a single enormous chip from a complete wafer, avoiding the traditional step of slicing it into smaller units. While specific transistor counts and process nodes are not detailed in the provided text, the physical scale is described as dramatic compared to standard packaging methods. The wafer-scale integration allows for massive internal bandwidth, reducing the latency typically associated with moving data between separate chips in a cluster.
Where it went
The WSE-2 powers the Cerebras CS-2 system, which is designed for high-end artificial intelligence problems. It supports scaling out to clusters of up to 192 units, effectively acting like a collection of rack-sized GPU clusters. The hardware is suited for training large neural networks where traditional graphics processing units might face memory or interconnect limitations. It handles tasks that benefit from algorithmic training improvements through sparsity and high utilisation of compute resources.
Versions and family
The WSE-2 succeeds the original Wafer Scale Engine. It works in conjunction with two key external technologies: MemoryX and SwarmX. MemoryX provides additional storage capacity ranging from 4TB to 2.4PB, using a mix of dynamic RAM and flash storage. SwarmX serves as the high-speed interconnect fabric, linking multiple CS-2 systems together with over 1Tb of bandwidth to facilitate weight streaming across the cluster.
Why it matters
This approach challenges conventional wisdom by scaling up physically rather than just packing more transistors into small dies. It demonstrates that wafer-scale engines can achieve near-linear performance scaling when solving large problems. By focusing on sparsity and memory expansion, Cerebras aims to offer a tenfold opportunity for performance gains without requiring proportional increases in silicon area, influencing how large models are trained in data centres.
Questions people ask about Cerebras WSE-2
How is the WSE-2 different from normal chips?
Standard chips are cut from a wafer into small individual pieces. The WSE-2 uses the entire wafer as one single connected processor, allowing for much higher internal bandwidth and memory capacity than traditional packaged dies.
What is MemoryX and why is it needed?
MemoryX is an external storage system that expands capacity beyond the 40GB of onboard SRAM. It holds model weights that do not fit on the chip itself, streaming them in as needed to support very large neural network layers.
What does sparse compute mean in this context?
Sparse compute allows the processor to ignore zero values in data calculations. This reduces the amount of work required, offering significant speedups and efficiency improvements while maintaining model accuracy during AI training tasks.
Good reading and sources
There is no English Wikipedia article on Cerebras WSE-2 yet, so the facts on this page come from the sources listed above: servethehome.com/cerebras-wafer-scale-engine-2-wse-2-at-hot-chips-33/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.