2
\$\begingroup\$

Limiting this to the current generation of consumer NPU hardware going into new PCs:

Can the NPU accelerator hardware being put in recent PCs practically be repurposed for non-AI tasks? And what tasks would it be good for if so?

I'm assuming someone might need to develop "GPNPU" APIs to use it. Given what little I know, I'm guessing maybe it could be used for convolutions like image sharpening or DSP filters for audio or SDR?

Or is the hardware too purpose built, like trying to compute on the raster-hardware of pre-shader 3d accelerators? Or is it maybe just hamstrung by weird AI number formats?

Since this is hardware design limited I'm assuming EE's the right site since I don't see a computer engineering exchange.

\$\endgroup\$
2
  • \$\begingroup\$ Since it is clearly a question about running different kinds of software on existing hardware, this would be suitable for software engineering. \$\endgroup\$
    – Dave Tweed
    Commented Aug 15 at 18:54
  • \$\begingroup\$ @DaveTweed It feels like SE-SE's topics are more high-level about programming itself, while this is almost more like numerical analysis? But I'll look for similar topics. \$\endgroup\$
    – davolfman
    Commented Aug 15 at 19:54

1 Answer 1

Reset to default
2
\$\begingroup\$

Can the NPU accelerator hardware being put in recent PCs practically be repurposed for non-AI tasks?

Depends on how the NPU has been implemented. Some NPU's are nothing more than repurposed DSP's so if that is the case you could use it for a DSP. Some are more specialized and only do multiply adds at specific bit widths, in these cases you could do multiply adds with them, but not very useful outside of AI.

\$\endgroup\$
5
  • 1
    \$\begingroup\$ Aren't multiply-adds also the main operation of signal processing, like FIR filters? \$\endgroup\$
    – davolfman
    Commented Aug 15 at 19:51
  • \$\begingroup\$ Yes, but not in the same manner as FIR because FIR can have different structures. You also need to have a lot of bit depth to avoid quantization noise in digital filters. NPU's may not be built to do that as AI does not need to be accurate, you can truncate or approximate many applications and they still work just not as good. \$\endgroup\$
    – Voltage Spike ♦
    Commented Aug 15 at 19:53
  • \$\begingroup\$ I guess that files under "Weird number formats". So the more AI-specific an accelerator is the more likely it is to use something like bfloat16 that's low-precision garbage you wouldn't want to use for other applications. So maybe it could do something quality-non-critical like real-time screen sharpening or something. \$\endgroup\$
    – davolfman
    Commented Aug 15 at 20:01
  • \$\begingroup\$ A lot of AI applications use int16's or int8's especially edge applications \$\endgroup\$
    – Voltage Spike ♦
    Commented Aug 15 at 20:29
  • \$\begingroup\$ If you give someone something that does tens of TOps of int16 MACs and I bet they will find a way to use it. On the other hand if the API's are so high level that drivers for a DSP and a specialized NPU can both implement it, there's little to no hope of getting one API for general compute: the hardware is just not even implementing the same concepts like shaders. \$\endgroup\$
    – davolfman
    Commented Aug 16 at 18:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.