Re: Peculiar Win/11 Behavior?
"Yuri Davidovsky (as work at disclosure dot ie)" <[email protected]>
| Newsgroups | gmane.lisp.lispworks.general |
|---|---|
| Message-ID | <[email protected]> |
> On 18 Jan 2025, at 23:57, David McClain <[email protected]> wrote: > > I think a better approach would be to couple the DSP as a specialized co-processor to a conventional CPU. It kinda works like that nowadays with gpu’s. They are not necessarily dsp specialised, but number crunching specialised, however in many cases it can mean the same. A RTX 4080 card has something like 40 teraflops of performance and the cuda sdk includes fft primitives. There is so much computing power in the modern gpus that keeping them fed can be a serious challenge. You might not even bother with fft at all and just dft yourself into oblivion and you won’t even max it out on the way there. On macos you have metal, apple’s gpu programming interface. What makes apple’s gpu programming different is that the cpu and gpu share the memory space and on M1 both come as a single chip, that helps with data transfer latency and data duplication between the devices you have to deal with on discrete gpus. Additionally m1 has a matrix multiplication co processor, as well as a neural engine one (should be good for dot products). I think there is also one for cosine transform acceleration, good for jpegs and all that. Those are not very well documented, however. _______________________________________________ Lisp Hug - the mailing list for LispWorks users [email protected] http://www.lispworks.com/support/lisp-hug.html