WSE (Wafer-Scale Engine) Architecture
What is WSE
WSE (Wafer-Scale Engine) is Cerebras Systems' wafer-scale AI processor, using an entire 12-inch wafer as a single chip (vs traditional GPUs occupying only a small piece of the wafer).
Cerebras WSE-3 (announced 2024-04) features 4 trillion transistors, 900,000 cores, 44 GB SRAM, purpose-built for ultra-large-scale AI training (Llama 3 405B, GPT-4 class).
WSE Core Innovation
Wafer-Scale Integration
- Traditional GPU chips ~800 mm² (~3% wafer area)
- WSE entire wafer = 46,225 mm² (57× larger)
- On-chip SRAM 44GB (vs H100 80GB HBM, but SRAM is 1000× faster than HBM)
- On-chip interconnect 214 PB/s (Fabric bandwidth)
Sparse Linear Algebra Cores (SLAK)
- Optimized for GEMM (matrix multiply) and sparse operations
- Faster than GPU tensor cores (specific workloads)
SwarmX Interconnect
- Multi-WSE systems: Cerebras CS-3 clusters
- Connected via SwarmX fabric (~1.2 TB/s)
- CS-3 Cluster = 2,048 WSE = 8 ExaFLOPS
Mainstream WSE Comparison
| WSE | Year | Transistors | Cores | SRAM | Process | Compute (FP16) |
|---|---|---|---|---|---|---|
| WSE-1 | 2019 | 1.2T | 400K | 18 GB | TSMC 16nm | - |
| WSE-2 | 2021 | 2.6T | 850K | 40 GB | TSMC 7nm | ~62 PFLOPS |
| WSE-3 | 2024 | 4T | 900K | 44 GB | TSMC 5nm | 125 PFLOPS |
| WSE-4 (CS-4, estimated) | 2027 expected | ~5T | ~1.5M | ~80 GB | TSMC 3nm | ~200 PFLOPS |
⚠️ WSE-4 is not officially announced; specifications above are estimates. Cerebras filed for IPO on 2026-04-17; WSE-4 will be the first public product after IPO.
WSE vs GPU
| Dimension | WSE-3 | H100 SXM | MI300X |
|---|---|---|---|
| Transistors | 4 trillion | 80 billion | 153 billion |
| Cores | 900K (sparse cores) | 14,592 CUDA | 14,592 SP |
| Memory | 44 GB SRAM | 80GB HBM3 | 192GB HBM3 |
| Memory bandwidth | ~1 PB/s | 3.35 TB/s | 5.3 TB/s |
| TDP | ~15 kW | 700W | 750W |
| Compute (FP16) | 125 PFLOPS | 989 TFLOPS | 1.5 PFLOPS |
| Deployment | Full system (Cerebras CS-3) | PCIe/SXM card | PCIe card |
| Best for | Ultra-large model training | General training | General training |
WSE Use Cases
- ✅ Ultra-large LLM training (Llama 3 405B, GPT-4 class)
- ✅ Genomics (biomedical)
- ✅ Scientific computing (HPC)
- ✅ Long sequence training (44GB SRAM enables massive batches)
- ❌ General AI inference (use GPU)
- ❌ Edge deployment (15kW TDP)
- ❌ Small/medium model training (cost-inefficient)
Commercial Deployments
- G42 (UAE AI company, $900M order)
- Mayo Clinic (medical AI)
- Argonne National Lab (scientific computing)
- Llama 3 405B training (Meta collaboration with Cerebras)
2026 Cerebras IPO (Major Event)
| Item | Details |
|---|---|
| IPO filing date | 2026-04-17 (S-1 filed) |
| Target listing date | 2026-05 (Nasdaq: CBRS) |
| Valuation | $22-25B |
| 2025 revenue | ~$510M (YoY +150%) |
| 2025 net loss | ~$200M (still losing, but narrowing) |
| Key deal | OpenAI $10B inference compute long-term contract |
| Major customers | OpenAI, G42, Mistral, Meta, Mayo Clinic |
| Underwriters | Goldman Sachs / Morgan Stanley / J.P. Morgan |
💡 IPO strategic significance:
- Cerebras is the world's second-largest wafer-scale AI company (behind NVIDIA by market cap)
- OpenAI $10B contract = largest single-customer inference compute order in history
- WSE-4 (CS-4) will be the first public post-IPO product (expected 2027)
- Competing with NVIDIA Groq 3 LPX for the ultra-low-latency inference market
Detailed Product Pages
- Cerebras WSE-2 - 2021 second generation
- Cerebras WSE-3 - 2024 third generation
- Cerebras WSE-4 - 2027 expected (first post-IPO generation)
Related Architectures
- GPU - General AI
- TPU - Google data center
- RPU/RDU - Reconfigurable architecture
- Complete Comparison Table