Cerebras wafer-scale AI
Cerebras: Wafer-Scale Systems for AI
Cerebras builds AI chips the size of a whole silicon wafer. Here is how a wafer-scale engine differs from a GPU, what the WSE-3 and CS-3 are, how they are delivered, and which claims to test before you plan around them.
Export controls: AI chips, servers, networking and memory can need an export licence depending on destination, end user and end use under US, UK and EU rules. We screen every project for this. See what applies.
How wafer-scale differs from a GPU
- The problem it targets. Conventional chips are limited to about 800 mm2 by the reticle used to print them, so large systems link many separate chips. Cerebras instead uses nearly an entire 300 mm wafer as one chip.
- Why that helps. Keeping compute, memory and fabric on one piece of silicon reduces the communication between chips that slows distributed workloads. Cerebras claims this makes scaling a very large model as straightforward as a small one.
- Handling defects. Cerebras designs for defects with redundant cores and routing, so a flawed region does not scrap the wafer.
- Our view of the trade-offs. On-chip SRAM is very fast but small, so very large models rely on external memory and streaming. That is a different design point from HBM-based GPUs, and it favours some workloads, notably fast inference, more than others.
The products
- WSE-3 (March 2024). 4 trillion transistors, 900,000 AI cores, 44 GB of on-chip SRAM and 125 PFLOPS of peak AI compute on TSMC 5 nm, in a 46,225 mm2 die.
- CS-3. The system around the WSE-3. External memory options range from 1.5 TB to 1.2 PB, enabling up to 24 trillion parameters in one logical memory space, and clusters can reach 2,048 systems.
- Condor Galaxy 3. A 64-system CS-3 cluster of about 8 exaflops owned by G42 in Dallas, as an example of a large deployment.
- Newer generation. Cerebras’s product page now lists a WSE-3T at 250 PFLOPS and a CS-4. We have not seen independent specifications for these, so we treat them as vendor claims.
Side by side
| Typical GPUReticle-limited die | WSE-3Wafer-scale chip | CS-3 systemThe system | ClusterScale-out | |
|---|---|---|---|---|
| Size | About 800 mm2 per die | 46,225 mm2, nearly a full 300 mm wafer | Packaged WSE-3 with external memory | Up to 2,048 CS-3 systems |
| Compute | Thousands of cores | 900,000 AI cores, 4 trillion transistors, TSMC 5 nm | 125 PFLOPS peak AI per chip (2024) | Condor Galaxy 3: 64 CS-3, about 8 exaflops |
| Memory | HBM beside the die | 44 GB of on-chip SRAM | External memory of 1.5 TB, 12 TB or 1.2 PB | Up to 24 trillion parameters in one logical space |
| Delivery | Servers and racks | Not sold alone | On-premises or Cerebras Cloud | Large customer builds |
Cerebras claims such as “19x more transistors” or “up to 30x faster inference than GPU systems” depend on workload and benchmark and are not verified here. Test your own models.
Software
- PyTorch. The Cerebras Software Framework offers native PyTorch 2.0 support, with a claim of much less code than GPU approaches for large models.
- Cloud and on-premises. You can run on Cerebras Cloud or buy systems for your own data centre.
- What it means for you. Check that your model architecture, precision and serving needs run on Cerebras’s compiler before you commit. See the platform comparison.
- Keeping it secure. Compilers, runtimes and the control plane need patching and review under our agent security and defence practice.
What it means for your data centre
- Power and cooling. CS-3 stays in the same power envelope as the previous generation but is a dense, specialised system. Cerebras’s datasheet covers power and cooling, and we confirm it for your site.
- Where it could fit. Fast inference for supported models, research on very large models, and teams that value a simpler programming model over a GPU ecosystem.
- Capacity options. Renting through GPU cloud providers or Cerebras Cloud can test the fit before you buy.
Where this fits in the stack
Common questions
What is a wafer-scale engine?
A single chip made from nearly a whole silicon wafer, with hundreds of thousands of cores and large on-chip memory, instead of many separate dies linked together.
Can I buy a CS-3?
Cerebras sells systems for on-premises use and offers Cerebras Cloud. Lead times and pricing come from Cerebras.
Is it faster than a GPU?
For some workloads, Cerebras claims large speedups, but results depend on the model and benchmark. Test your own workload.
Do export controls apply?
Yes. Advanced AI systems can need a licence depending on destination, end user and end use. We screen every project.
Sources and dating
Figures are as of 18 November 2026, from Cerebras’s WSE-3 announcement (March 2024), IEEE Spectrum and Cerebras’s chip page. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against current datasheets when we specify a build. Cerebras, WSE and CS-3 are trademarks of Cerebras Systems. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: Other Chip Providers · NVIDIA Data Centre GPUs · AMD Instinct GPUs · GPU Cloud · Hosted AI Models
Considering wafer-scale
Test it against your workload
