Edited by 
Short answer
The Cerebras WSE-3 (Wafer-Scale Engine 3) is an AI processor made from an entire silicon wafer instead of a small die. Announced in March 2024 and built on TSMC's 5 nm process, it has 4 trillion transistors, 900,000 AI cores and 44 GB of on-chip SRAM with 21 PB/s of memory bandwidth, and Cerebras rates it at 125 petaflops of peak AI compute. Keeping the model's data on the chip removes most of the memory bottleneck that limits GPUs, which is why Cerebras systems are known for extremely fast LLM token generation.
What is a wafer-scale chip?
Chips are made on round silicon wafers about 300 mm across. Normally, a wafer is cut into dozens or hundreds of individual dies, each becoming a separate processor. Even the largest GPUs, such as NVIDIA's B200, are limited to around 800 square millimetres per die by the lithography equipment, so big AI systems connect thousands of separate chips with networking.
Cerebras takes the opposite approach: it keeps the whole usable square of the wafer as one processor. The WSE-3 measures 46,225 mm², about 57 times larger than a flagship GPU die. Instead of moving data between chips over cables, data moves across the wafer on an on-chip network.
WSE-3 specifications
| Specification | Cerebras WSE-3 |
|---|---|
| Announced | March 2024 |
| Process | TSMC 5 nm |
| Chip area | 46,225 mm² |
| Transistors | 4 trillion |
| AI cores | 900,000 |
| On-chip memory | 44 GB SRAM |
| Memory bandwidth | 21 PB/s |
| Peak AI compute | 125 petaflops (Cerebras rating) |
| System | Cerebras CS-3 |
For comparison, NVIDIA's B200 has 208 billion transistors across two dies and 192 GB of HBM3E at 8 TB/s. The WSE-3 has far less memory capacity but around 2,600 times more memory bandwidth, because its memory is SRAM spread right next to the cores.
How the WSE-3 works
Cores with their own memory
Each of the 900,000 cores has its own small block of SRAM. This means data does not have to travel to a separate memory chip; it sits beside the core that uses it. SRAM is the same fast memory used for caches inside CPUs and GPUs, but on the WSE-3 it is the main memory.
A fabric instead of a network
The cores are connected in a two-dimensional mesh called the fabric. Data flows between neighbouring cores in a single clock cycle, so moving activations from one layer to the next is fast and cheap in energy.
Dealing with defects
Every wafer has manufacturing defects, which is why chipmakers normally use small dies: a defect ruins only one die. Cerebras builds in spare cores and redundant links, and routes around faulty cores, so the wafer still works as one chip.
Weight streaming and MemoryX
44 GB is not enough to hold the largest models' weights and optimizer state for training. For training, Cerebras uses weight streaming: weights live in an external memory system called MemoryX and stream onto the wafer layer by layer, while activations stay on chip. Up to 2,048 CS-3 systems can be clustered for very large models.
Why Cerebras is so fast at inference
As explained in our guide to GPU memory bandwidth, generating text with an LLM is limited by how fast weights can be read for each token. GPUs read them from HBM at a few terabytes per second. The WSE-3 keeps weights in SRAM at petabytes per second, so token generation for a single user can be many times faster.
When Cerebras launched its inference service in August 2024, it reported about 1,800 tokens per second for Llama 3.1 8B and about 450 tokens per second for Llama 3.1 70B, far beyond typical GPU cloud speeds for single requests at the time. For larger models, the weights are split across several wafers. Fast generation is especially valuable for reasoning models and AI agents, which produce many tokens before giving an answer.
Cerebras vs NVIDIA GPUs
| Cerebras CS-3 (WSE-3) | NVIDIA GPU systems | |
|---|---|---|
| Building block | One wafer-scale chip per system | Many GPUs linked by NVLink and networking |
| Memory | 44 GB on-chip SRAM, very high bandwidth | Large HBM capacity per GPU, lower bandwidth |
| Strength | Very fast single-stream inference, simpler scaling | Flexibility, huge software ecosystem (CUDA) |
| Software | Cerebras SDK, PyTorch support | CUDA, PyTorch, TensorRT and almost every AI tool |
| Typical buyers | Cloud inference, research labs, national labs | Nearly everyone |
GPUs remain the default because of CUDA's ecosystem, their flexibility and the huge memory capacity of systems such as the GB200 NVL72. Cerebras competes on speed and simplicity for specific workloads, especially inference. It belongs to a group of alternative AI chip designs alongside Groq's LPU, which also relies on on-chip SRAM; see Groq LPU vs GPU.
Power and cooling
A CS-3 system draws on the order of 20 kilowatts and uses liquid cooling, because the wafer packs so much compute into one place. Per system that is high, but it replaces many GPUs and their networking, so Cerebras argues the efficiency per unit of work is competitive.
Frequently asked questions
How big is the Cerebras WSE-3?
It is 46,225 mm², roughly 21.5 cm on each side, about the size of a dinner plate.
Is Cerebras faster than NVIDIA?
For single-user LLM token generation, Cerebras systems are typically much faster. For high-throughput batch serving, training at the largest scales and general flexibility, NVIDIA GPU clusters remain the standard choice.
Can I buy a Cerebras chip?
Cerebras sells complete CS-3 systems and clusters to organisations and offers cloud access to its inference service. It does not sell individual chips.
Sources and further reading
- Cerebras, Cerebras announces third-generation wafer-scale engine (March 2024)
- Cerebras, Introducing Cerebras Inference (August 2024)
- NVIDIA, Blackwell architecture
Leave a Reply