Skip to main content

Kernel Development

Here are examples of Kernel development

Basic Usage​

Step 1: DRAM Loopback​

Learn the basic structure of a Metalium application by implementing a DRAM loopback that copies data from one DRAM buffer to another.

What you’ll learn:

  • Basic host and kernel structure
  • Buffer management
  • Data transfer

Step 2: Eltwise Binary Kernel​

Build on the loopback example by implementing an Elwise binary kernel that performs addition on two buffers. This introduces computation using the matrix engine (FPU) and data passing within a Tensix core.

What you’ll learn:

  • Circular buffers for data passing
  • Compute kernels
  • Using the matrix engine (FPU)

Step 3: Eltwise SFPU​

Extend the previous example to implement an Eltwise SFPU kernel that performs element-wise addition using the SFPU (vector engine, Special Function Processing Unit). This will introduce you to the SFPU and how to use it for vectorized operations

What you’ll learn:

  • Vectorized operations using the SFPU

Intermediate Usage​

Step 4: Single-core Matrix Multiplication​

Implement a matrix multiplication on a single Tensix core using the matrix engine.

What you’ll learn:

  • Complex dataflow
  • Tiled operations
  • Matrix engine utilization

Advanced Usage​

Step 5: Multi-core Matrix Multiplication​

Extend to a multi-core implementation by distributing computation across multiple Tensix cores.

What you’ll learn:

  • Parallel processing
  • Workload distribution across cores

Step 6: Optimized Multi-core Matrix Multiplication​

Optimize the multi-core implementationby leveraging the processor grid, reducing redundant DRAM access, and minimizing NoC congestion through data sharing.

What you’ll learn:

  • Performance optimization techniques
  • Efficient data movement and reuse

Reference: https://docs.tenstorrent.com/tt-metal/latest/tt-metalium/get_started/get_started.html