Tenstorrent Grayskull / Wormhole
Product Overview
Tenstorrent, led by legendary architect Jim Keller, develops AI accelerators based on the RISC-V architecture. The Tensix core integrates a RISC-V CPU, matrix/vector/data transfer engines. Wormhole connects via standard Ethernet RDMA (no NVLink lock-in), and the software stack — TT-Metalium and TT-BUDA — is fully open-source.
Core Specifications
| Item | Grayskull (e150/e300) | Wormhole (n150/n300) |
|---|
| Architecture | Tensix | Tensix |
| Process | TSMC 12nm | TSMC 12nm |
| Tensix Cores | 120 | 144 |
| On-chip SRAM | 120 MB | 144 MB |
| FP32 | ~50 TFLOPS | ~75 TFLOPS |
| INT8 | ~300 TOPS | ~450 TOPS |
| Interconnect | PCIe 4.0 | 8× 100GbE RDMA |
| TDP | 75W / 300W | 150W / 300W |
Manufacturer Info
Key Features
- Tensix Architecture: RISC-V + Matrix + Vector + Data Transfer
- No HBM: fully on-chip SRAM (suited for specific models)
- Ethernet Interconnect: avoids NVLink vendor lock-in
- Fully Open-Source Software Stack
- Next-gen Blackhole: 2025 release, 2× performance
Use Cases
- Open-source AI developers
- Medium-scale AI training
- Academic research
- RISC-V ecosystem