Is Asteroids@home thinking of implementing special apps for NPU support for newer CPU processors which will have NPU included for AI calcs?

& will there be additional NPU support also for additional devices, such as Google TPU device Coral?

Thanks for answering,

---

non-profit org. Play4Life in Zagreb, Croatia, EU

you keep asking this at different projects. do you understand what an NPU is and does? if you did, you would probably realize that they are not suited for just about any BOINC project.

NPUs are used for low precision inferencing on AI tasks. it's kind of like asking if you can use your bitcoin miner for BOINC. they are both ASICs designed for one thing, and unless you're doing that one thing, they wont be able to be used.

the only exception would be if these integrated NPUs had built in logic in their task scheduler to recognize matching calls and instructions that could be offloaded to the NPU. but in that case, it would all be internal to the device itself, and it would be totally transparent to the project and the users with no changes needed. but since the vast majority of the work here is the double data type (FP64), the NPU would have little or nothing to do.

the only project that MIGHT have any hope to use an NPU could be LLMentor-Grid, and only if they are using or plan to use low precision (4 or 8 bit) in their computations.

---

The primary goal of an NPU is power efficiency. It functions similarly to a video accelerator in a GPU, which excels at decoding video efficiently but it's unusable for anything else.
You can decode a video on a general-purpose CPU or a GPU, but the efficiency won't be as good and it'll drain your battery faster. That’s also why desktop PCs don’t include an NPU, they don’t need one.

NPUs also don't have a standardized API like an OpenCL Each vendor has their specific API, so it's hard to utilize them reliably on a platform like BOINC.

New projects might find uses for an NPU in AI tasks, but it won't be technically usable for most current BOINC apps, even if developers wanted to.
Remember that GPUs will continue to be significantly more powerful for AI and easier to work with, which is another reason I don’t expect any BOINC NPU projects in the near future.

Thanks for the detailed technical explanations, Ian&Steve and ahorek. I completely agree that the current NPU architecture is fundamentally unsuited for the FP64 heavy lifting required by BOINC projects.

It's not about forcing the NPU to run the actual science compute, but rather exploring its potential for system-level power management. On modern heterogeneous architectures like the Snapdragon X2 Elite, the NPU excels at offloading background AI-driven OS tasks. By delegating those minor inference workloads to the NPU, we can keep the CPU cores clean and dedicated 100% to BOINC crunching with minimal context switching and thermal noise.

It is a niche perspective, but as we move toward these highly integrated SoCs, optimizing the entire power-per-watt profile—rather than just the compute core—could be a very interesting future direction for BOINC. Even if it doesn't change current application requirements, it’s a fascinating look at how these platforms might evolve.

copilot@home :)