Skip to main content

TT-Metal Overview

TT-Metal is a low-level Software Development Kit (SDK) for Tenstorrent Neural Processing Units (NPUs), providing direct access to the underlying hardware. It consists of two core layers—TT-NN and TT-Metalium—and enables developers to fully harness the performance of Tenstorrent chips. As the foundation of the entire software stack, all higher-level compilers and runtimes—including TT-Forge and TT-NN—ultimately operate through TT-Metal.

  • TT-NN : An optimized neural network operator library with a PyTorch-like API
  • TT-Metalium : A low-level programming model for developing and optimizing custom kernels
note

TT-Metal is designed for developers who want to work at the closest level to the hardware.

alt text


Key Features​

TT-Metal offers various development options at different levels:

Performance Optimization​

  • Support for various AI models
  • Optimized operation libraries
  • Data format and memory management options

Advanced Development​

  • Custom kernel development
  • Hardware-level optimization
  • C++ programming support

Relationship Between TT-Metal and TT-Forge​

TT-Metal sits at the lower level of the Tenstorrent software stack, while TT-Forge is at the higher level:

  • TT-Forge: A compiler that transforms models from frameworks like PyTorch, TensorFlow, JAX, and ONNX to run on Tenstorrent hardware.
  • TT-Metal: A low-level SDK that provides direct hardware access for maximum performance.

When to Use TT-Metal​

TT-Metal is recommended in the following scenarios:

  • When you need maximum performance at the hardware level
  • When developing custom AI models or operations
  • When optimizing for specific AI workloads

Getting Started​

To get started with TT-Metal: