Parts/Accelerators/Cerebras WSE-3
Accelerators · Wafer Scale Engine · 2024
Cerebras WSE-3
A single processor occupying an entire silicon wafer, roughly fifty times the area of a reticle-limited die. The design removes the off-chip interconnect for models that fit on it.
Why this part
One chip the size of a dinner plate, because the reticle limit is a manufacturing constraint rather than a law: Cerebras stitches across it and keeps the entire wafer as a single device, then routes around the defects that would normally make that impossible. 900,000 cores and 44GB of on-wafer SRAM at 21 PB/s means no off-chip memory in the training loop at all. It draws 23 kilowatts and needs its own cooling and power design, and the pitch is that one of these replaces a cluster along with the network and the parallelism work that a cluster requires.
Systems using this part
- Systems
- 4
- Rmax underneath
- 0.0 GFlop/s
- Units recorded
- 366
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| Cerebras Oklahoma City 2025-09 | Accelerator | USA | 300 | - | Cerebras | - |
| Neocortex 2021-03 | Accelerator | USA | 2 | - | Cerebras | - |
| Condor Galaxy 3 | Accelerator | USA | 64 | - | Cerebras | - |
| Cerebras Guyana AI Data Center | Accelerator | Guyana | - | - | Cerebras | - |
“Units recorded” sums only the edges where a public source states a quantity. It is a floor, not a total, and should never be read as installed base.