tools
TT-Metalium
Open-source low-level SDK for writing custom C++ kernels and applications on Tenstorrent hardware.
TT-Metalium gives developers direct control over Tensix hardware, including RISC-V processors, the Network-on-Chip, and matrix and vector engines. It provides C++ APIs for writing, configuring, and executing kernels, with explicit control over memory and data movement.
It is intended for developers building custom machine-learning or high-performance-computing workloads, including novel operations and distributed multi-chip programs. It requires compatible Tenstorrent hardware and software setup; no public software pricing was found.
Features
- C++ API for writing and executing kernels on Tensix hardware
- Direct access to RISC-V processors and the Network-on-Chip
- Control over matrix and vector engines
- Explicit memory allocation and data movement
- Custom single-core and multi-core kernel development
- Support for multi-device and distributed computations
- Open-source code under the Apache-2.0 license
Use cases
- Write custom kernels for machine-learning operations
- Optimize memory movement and compute across Tensix cores
- Build high-performance computing applications
- Implement operations not available in standard libraries
- Port workloads from one Tenstorrent device mesh to another
Pros
Cons
Latest updates
- v0.83.0-dev20261105 (v0.83.0-dev20261105)
top32_rm sizes the L1 chunk offset by the source format instead of the Dest format.
- v0.82.0-dev20261104 (v0.82.0-dev20261104)
Fixes a matmul deadlock with partial face tiles.
- nomic-v03: nomic: MoE dispatch and combine on the generated ops (nomic-v03)
MoE dispatch and combine run on generated ops for windows of 1000 to 4096 tokens.
- nomic-v02: nomic: residual stream on the generated layer norm's ND layout (nomic-v02)
The residual stream uses the generated layer norm's ND layout.
- nomic-v01: nomic: fused attention front (nomic_qkv_rotary_heads) on today's layout (nomic-v01)
QKV, head split, scaling, and rotary run as one generated fused attention op when the layout and buffers allow.
Capabilities
- API — “TT-Metalium provides a C++ API for writing kernels that run on Tensix hardware; for configuring these kernels; and for executing kernels on the hardware.” source
- Official SDKs — “TT-Metalium™ is Tenstorrent’s open source, low level AI hardware SDK, getting you as close to the metal as possible for custom kernel development, experimentation, and more.” source
Get it
Pricing
- Starting price
- Free
- Prices checked
- 2026-10-25
TT-Metalium
Free
- Open source low-level AI hardware SDK
- C++ API for custom kernels on Tensix processors
- Direct access to RISC-V, NoC, Matrix/Vector engines
- Available on GitHub under open-source license
- No licensing fees