Running custom C kernels on Intel NPU for unsupported inference tasks
Problem
Developers want to run custom inference workloads on Intel NPU that are not officially supported, but existing toolchains like OpenVINO and llama.cpp have limited support, requiring manual kernel development and dealing with platform compatibility and support issues.
Pain points
- Lack of official support for custom NPU inference tasks
- Difficulty using NPU with OpenVINO and llama.cpp
- Demand for broader DSP support such as Hexagon
Questions to verify
- Do you have specific inference tasks that need to run on the NPU?
- Are you willing to try using it on Linux and contribute code?
- How interested are you in support for Hexagon or other DSPs?
Evidence
quote · @Anonymous
Finally, bare-metal NPU programming! Immediately thought of those custom, weird inference tasks Intel won't ever officially support.
fact
Requires Windows x64plonk.
Disclaimer: this card is distilled by AI from community discussions as “pain point evidence”. It is a starting point for need analysis and validation, not a “validated market need”. AI has analytical limits; please open the original discussion above to manually verify before key decisions.