TT-Metalium

TT-Metalium is Tenstorrent’s open-source, low-level AI hardware SDK that provides a C++ API for writing kernels targeting Tensix processors, giving developers direct access to RISC-V cores, the Network-on-Chip (NoC), and Matrix/Vector engines. It is designed for engineers who need full control over hardware and data movement — enabling custom kernel development, experimentation, and optimization for both ML and non-ML workloads. The SDK sits at the foundation of Tenstorrent’s software stack, below higher-level libraries like TTNN (which offers Python APIs for common ML operations) and TT-Forge/TT-MLIR (compilation frameworks for neural network deployment). It is best suited for advanced users comfortable with low-level programming and hardware architecture, as it exposes every detail of the Tensix core for fine-grained tuning.

TT-Metalium follows a bottom-up programming philosophy: start with a single Tensix core kernel, then schedule across multiple cores with synchronization, and finally scale to multi-device deployments. Operations typically use a three-stage pipeline — a reader kernel moves data from DRAM/SRAM into circular buffers, a compute kernel processes data using Matrix/Vector engines, and a writer kernel writes results back. Circular buffers act as FIFO structures enabling overlapped execution (fetching new data while compute processes previous data). The SDK includes features like reader, compute, and writer kernel abstractions, and supports custom fused kernels and tensor manipulations (TMs). The GitHub repository (tt-metal) has over 1,500 stars, 484 forks, and 27,463 commits as of June 2026, reflecting active development.

In the low-level AI hardware SDK market, TT-Metalium competes directly with NVIDIA's CUDA (proprietary, closed-source) and AMD's ROCm (open-source but less mature for AI). Unlike CUDA, which encrypts APIs and hides hardware details, TT-Metalium is fully open-source with no black boxes or hidden functions — developers can inspect and modify every layer. It also differs from Intel's oneAPI (which abstracts across hardware) by providing direct, unabstracted access to Tenstorrent's Tensix architecture. For non-ML workloads, it competes with general-purpose GPU compute frameworks like OpenCL, but TT-Metalium is purpose-built for Tenstorrent's specific hardware topology. Its openness is a differentiator against proprietary SDKs, but its ecosystem and community size are smaller than CUDA's.

The honest trade-offs: TT-Metalium requires deep knowledge of Tensix hardware details (RISC-V cores, NoC topology, memory hierarchy) and low-level C++ programming, making it unsuitable for beginners or rapid prototyping. The SDK is complex — developers must manually manage data movement, synchronization, and kernel orchestration, which increases development time compared to higher-level frameworks like PyTorch. Documentation, while improving, is still evolving (the docs site has a single guide with limited examples), and community support is primarily through a Discord server rather than extensive forums or tutorials. Finally, TT-Metalium is tied exclusively to Tenstorrent hardware (Grayskull, Wormhole, etc.), so code written for it cannot run on NVIDIA, AMD, or Intel GPUs without a complete rewrite.

Open source (free, no licensing fees; available on GitHub under an open-source license)