But at a certain point I guess it just breaks? And they need an objective "I gave these tokens, I got out those tokens". But I guess that would need an objective gold standard ground truth that's maybe hard to come by.
But at a certain point I guess it just breaks? And they need an objective "I gave these tokens, I got out those tokens". But I guess that would need an objective gold standard ground truth that's maybe hard to come by.
If you're in the business of selling unbundled edge accelerators, you're strongly incentivized to modularize your NPU software stack for arbitrary hosts, which increases the likelihood that it actually works, and for more than one particular kernel.
If I had an embedded AI use case, this is something I'd look at hard.
This is an honest, neutral question, and it's specifically about what can concretely be done with them right now. Their theoretical use is clear to me. I'm explicitly asking only about their practical use, in the present time.
(One of the reasons I am asking is I am wondering if this is a classic case of the hardware running too far ahead of the actual needs and the result is hardware that badly mismatches the actual needs, e.g., an "NPU" that blazingly accelerates a 100 million parameter model because that was "large" when someone wrote the specs down, but is uselessly small in practice. Sometimes this sort of thing happens. However I'm still honestly interested just in what can be done with them right now.)
I couldn't imagine recommending any of these boards to people who aren't already SBC tinkerers.
There are some perplexity comparison numbers for the previous gen - Orange pi 5 in link below.
Bit of a mixed bag, but doesn't seem catastrophic across the board. Some models are showing minimal perplexity loss at Q8...
https://github.com/invisiofficial/rk-llama.cpp/blob/rknpu2/g...
Is this a thing? I read an article about how due to some implementation detail of GPUs, you don't actually get deterministic outputs even with temp 0.
But I don't understand that, and haven't experimented with it myself.
The main difference comes from rounding order of reduction difference.
It does make a small difference. Unless you have an unstable floating point algorithm, but if you have an unstable floating point algorithm on a GPU at low precision you were doomed from the start.