← All hardware
A single wafer-scale chip the size of a dinner plate for record inference speed.
Pros
- Fastest measured LLM inference token throughput
- Eliminates multi-GPU model-sharding complexity
- Huge on-chip memory bandwidth (21 PB/s)
Cons
- Very high power and physical footprint (~23 kW per CS-3)
- Niche, specialized ecosystem vs GPUs
- Limited on-chip memory (44 GB) for largest weights
✓ Where it shines / best for
- Ultra-fast LLM inference (token-rate-critical apps)
- Research labs training large models on few devices
- Organizations wanting an inference cloud API
✕ Not the best fit for
- Standard rack/power-constrained data centers
- Buyers needing a broad GPU software ecosystem
- Small budgets (system-scale pricing)
Features
- ✓ Single 300mm TSMC 5nm wafer, 4 trillion transistors
- ✓ 900,000 AI-optimized cores
- ✓ 125 PFLOPS peak AI performance
- ✓ 44 GB on-chip SRAM at 21 PB/s bandwidth
- ✓ No multi-GPU partitioning of models needed
- ✓ Industry-leading LLM inference token rates
- ✓ Cloud inference API and on-prem CS-3 systems
- ✓ Simplified single-device programming model
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| Cloud inference - Free tier | $0 | daily free tokens | Free API access with generous daily limits (e.g. ~1M free tokens/day on login) |
| Cloud inference - Developer (pay-as-you-go) | $0.10-$0.60+ per 1M tokens | per million tokens | Per-token pricing varies by model; e.g. Llama 3.1 8B ~$0.10/M and Llama 3.1 70B ~$0.60/M, larger/newer models priced higher |
| Enterprise / CS-3 system (purchase) | Contact sales | one-time / custom | System-scale hardware pricing (historically several million USD per CS-3); managed private cloud or on-prem; custom SLAs. Contact Cerebras. |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| cores | 900,000 |
| form_factor | CS-3 system (~23 kW) |
| performance | 125 PFLOPS peak AI |
| transistors | 4 trillion |
| architecture | WSE-3 wafer-scale engine (TSMC 5nm) |
| on_chip_sram | 44 GB at 21 PB/s |
Sponsored
A full review is being generated for this product and will appear here shortly.