Tenstorrent AI Accelerator
Vendor: Tenstorrent
Category: ASIC Dedicated Accelerator
Architecture: Tensix / RISC-V
Introduction
Tenstorrent's AI accelerators (Wormhole/Grayskull/Blackhole) use a unique Tensix architecture combining RISC-V cores with AI tensor compute units. Led by legendary chip architect Jim Keller, with an open-source software stack.
Specifications
| Model | Compute | Memory | Interface | TDP | Process |
|---|---|---|---|---|---|
| Wormhole n300 | 200 TFLOPS (BF16) | 64GB LPDDR5 | PCIe 5.0 + Ethernet | 200W | 6nm |
| Grayskull e150 | 50 TFLOPS (BF16) | 16GB GDDR6 | PCIe 4.0 | 75W | 12nm |
Official Website
Driver Downloads
Linux
Related Documentation
- Tenstorrent Developer Documentation
- TT-Metalium Programming Guide
- TT-MLIR Compiler
- Wormhole Architecture Introduction
OS Support
| Windows | Linux | macOS | Android |
|---|---|---|---|
| ❌ | ✅ | ❌ | ❌ |
Version History
| Version | Release Date | Description |
|---|---|---|
| TT-Metalium 1.0 | 2024 | RISC-V open-source software stack |
Performance Benchmarks
| Model | Task | Performance Metric |
|---|---|---|
| Wormhole n300 | LLM Inference | Medium-scale inference optimized |
| Grayskull e150 | Edge AI | Entry-level |
Pricing
| Model | Reference Price | Notes |
|---|---|---|
| Wormhole n300 | ~$699-999 | Developer card |
| Grayskull e150 | ~$399-599 | Entry-level developer card |
Quick Installation
Linux (Ubuntu 22.04)
# 1. Install Tenstorrent driver
pip install tt-torch
# 2. Verify
tt-smi
Obtain drivers and tools from Tenstorrent GitHub.
Code Examples
Python (PyTorch TT)
import torch
import tt_lib
# Use Tenstorrent backend
device = tt_lib.device.CreateDevice()
tt_lib.device.SetDefaultDevice(device)
# tt-torch provides PyTorch backend
x = torch.randn(1024, 1024)
# Automatically routes to Tenstorrent hardware
Architecture Highlights
- Tensix Core: Programmable tensor cores based on RISC-V, each containing a matrix engine and routing engine
- Wormhole Interconnect: High-bandwidth inter-chip/card interconnect supporting mesh topology
- Open-Source Path: Open hardware design and software stack with an active developer community
Model Compatibility
| Model/Framework | Support | Notes |
|---|---|---|
| PyTorch | ✅ tt-torch | Backend under development |
| Cerebras NN | ✅ | Similar dataflow architecture |
| LLM Inference | ⚠️ | Community adaptation in progress |
| General AI | ⚠️ | Early ecosystem stage |
Related Products
If you're evaluating alternatives, the following products may also fit your scenario:
- Google Cloud TPU — Google (TPU Tensor Processor)
- Cerebras WSE-3 — Cerebras (ASIC Dedicated Accelerator)
- Etched Sohu ASIC — Etched (ASIC Dedicated Accelerator)
- NVIDIA GPU / CUDA — NVIDIA (GPU Graphics Processor)
- Intel Gaudi (Habana) — Intel (ASIC Dedicated Accelerator)
- SambaNova RDU — SambaNova (ASIC Dedicated Accelerator)
- AWS Trainium / Inferentia — Amazon AWS (ASIC Dedicated Accelerator)