MN-Core Software Samples & Build Resources

This repository provides the following resources for working with MN-Core:

• Examples for the MN-Core SDK
• Resources for setting up the MN-Core SDK environment
• Resources for managing bare-metal machines equipped with MN-Core

Try the MN-Core SDK

The MN-Core SDK enables developers to build and optimize programs for the MN-Core architecture. The SDK consists of two primary components:

• MLSDK: Provides PyTorch-compatible interface for developing machine learning models.
• HPCSDK: Provides a general-purpose programming environment in C/C++ with OpenCL-compatible and directive-based programming models. Currently, only MNCL, an OpenCL-compatible environment, is provided.

Programs developed with the MN-Core SDK are portable. They can run on MN-Core hardware and emulators, and MLSDK programs can also run on GPUs. This flexibility allows you to verify the compatibility of your programs quickly even without direct access to MN-Core hardware.

We recommend installing the MN-Core SDK environment using Docker for a consistent setup. Please refer to the following documentation for detailed installation instructions:

• sdk/0.5/README.md

When using MLSDK, the following documentation will be helpful:

• MLSDK Documentation
• MLSDK Examples

When testing HPCSDK, please refer to the header files included with the SDK. Note that HPCSDK is currently in an alpha stage.

Run Programs with MN-Core

Programs developed with the MN-Core SDK can run in the following three environments:

Run on the emulator

The MN-Core SDK includes an MN-Core emulator that runs programs without access to actual hardware. This allows you to easily evaluate portability to MN-Core and estimate expected performance.

The usage of the emulator differs between MLSDK and HPCSDK. Please refer to their respective documentation for details.

Run on bare metal (MN-Core 2 Devkit / MN-Server 2)

If you have access to an MN-Core 2 Devkit or MN-Server 2, you can run programs on actual MN-Core hardware instead of the emulator. These systems allow you to execute programs within a Docker container built using the instructions in sdk/0.5/README.md.

Run on the cloud (Preferred Computing Platform; PFCP)

The Preferred Computing Platform (PFCP) also provides environments equipped with MN-Core processors. You can use pre-built Docker images on PFCP that are equivalent to images built with sdk/0.5/README.md. This makes it straightforward to deploy MN-Core applications on PFCP after verifying them locally with the emulator or on bare-metal systems. For instructions on using MN-Core on PFCP, please refer to the PFCP documentation.

Repository Contents

• apt/ -- Resources to install MN-Core packages via apt
  • add_mncore_packages.sh -- Script to add the APT repository in your Ubuntu system
• sdk/ -- MN-Core SDK
  • 0.4/ -- MN-Core SDK 0.4
    • mncore-sdk-minimal.Dockerfile -- Dockerfile to build the MN-Core SDK 0.4 minimal Docker image with dependencies
    • mncore-sdk-full.Dockerfile -- Dockerfile to build the MN-Core SDK 0.4 "full" Docker image with extra packages
    • create_dev_ctr.sh -- Script to start an MN-Core SDK 0.4 development container
  • 0.5/ -- MN-Core SDK 0.5
    • mncore-sdk-minimal.Dockerfile -- Dockerfile to build the MN-Core SDK 0.5 minimal Docker image with dependencies
    • mncore-sdk-full.Dockerfile -- Dockerfile to build the MN-Core SDK 0.5 "full" Docker image with extra packages
    • create_dev_ctr.sh -- Script to start an MN-Core SDK 0.5 development container
  • examples/ -- Examples for the latest MLSDK