NPU/TPU support?
21 Feb 2025 6:11:03 UTC
Topic 232124
(moderation:
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,
KLiK wrote:Is
)
KLIK,
Some of your questions, like the ones you posted in this post, might be better answered by the developers of the projects related to your questions.
I might suggest you try out Discord. It can be found here: https://discord.com/download and is available for any device such as Linux and Windows, as well as others. You will find that on Discord there are the individual BOINC projects as well many others.
Proud member of the Old Farts Association
I am no developer, but this
)
I am no developer, but this would be very unlikely, because NPU (neural processing units) are not able to process 32 bit numbers. Only 4 bits. The processing and programming is totally different to CPU and GPU.
The same with TPU (Tensor processing units). They could only be programmed in CORAL.
With TPUs you compare pictures. Wikipedia, should have more information.
you asked this same question
)
you asked this same question like 6 months ago on several projects. the answer wont change since then.
NPUs are ASICs that do low precision inferencing. we do not do that here or at any project (*maybe* LLMentor-Grid could use it). and as ahorek commented elsewhere, the lack of a standardized API for these devices makes it very difficult to design for. and BOINC would not even recognize these devices.
_________________________________________________________________________
Ian&Steve C. wrote: you
)
Yes, I did about the TPU...as they were just getting used as add-in for laptops & USB ports for computers.
But, back then CPU's of Intel Core Ultra CPU (15th gen) were not developed & on market...now they are, so a question came here to be asked. As these things will become more common, so I ask again. Why? Sometimes the time is not right, but that does not seem that with a time some circumstances will not change - & the time would be right.
As for the device recognition, Windows does recognize it...so it is simple to include them.
non-profit org. Play4Life in Zagreb, Croatia, EU
A TPU is a type of NPU. The
)
A TPU is a type of NPU. The nature of the question doesn’t really change.
math projects won’t be able to use it either. Precision is too low. it’s an ASIC that can only be used for one type of calculation.
being detected by windows doesn’t matter. BOINC has no mechanism to detect it. It’s like when you have a GPU installed without the drivers. Windows can see the device, but BOINC is oblivious to it. BOINC has no idea how to interface with that kind of device. And none of them will interface in the same way (not standardized) so there’s likely not an easy way for BOINC to even try to add them without the tedious task of coding in all the possible ways that all the NPUs communicate.
It’s just not going to happen. NPUs are for AI inference. That’s it.
_________________________________________________________________________
Ian&Steve C. wrote: A TPU is
)
I agree to disagree, as there will be some NPU implementation in some of the project(s).
& again, if Windows see it, then it has drivers - otherwise it is "unknown device of a sort". Will BOINC going be implemented to see it, we will see. As it can check for additional hardware.
But my bet is that some projects will use the NPU processing within their app, for newer set of instruction within CPU architecture.
Now that it is written, we can wait & see. B)
non-profit org. Play4Life in Zagreb, Croatia, EU
Maybe for comparison of DNA
)
Maybe for comparison of DNA or similar things, NPU could become useful in the future, but for computing of data, it is not. It is a totally different thing. A NPU is no GPU.