By: Andrey (andrey.semashev.delete@this.gmail.com), April 10, 2021 10:36 am
Room: Moderated Discussions
Robert Williams (crispysilicon.delete@this.gmail.com) on April 10, 2021 8:24 am wrote:
> Mark Roulo (nothanks.delete@this.xxx.com) on April 9, 2021 12:54 pm wrote:
> > My employer runs a lot of hand optimized numeric code. We started with AltiVec
> > on PowerPC, then moved to SSE on x86 and are now using AVX2 on x86.
> >
> > One reason we HAVE NOT moved to Avx-512 is that the developers like being able to run code on their
> > desktop/laptop machines. Until recently, Avx-512 was pretty much only available on server parts.
> >
> > Making it difficult (or impossible) to run code on developer
> > machines discourages adoption of new capabilities.
> >
> > Note that a SLOW Avx-512 implementation on the laptop/desktop chips would have
> > been okay. The developers (other than the performance optimization) team don't
> > need the code to run QUICKLY on their machines -- they just need it to run.
>
> I get you, but it's not the same.
>
> In the case of AVX-512, it's going to see some use for average Joe. This isn't about development.
> This is about the point that for TSX to be fast enough to be worth using, it's going to require
> complexity that is going to be counterproductive to desktop/mobile. It's going to sit cold.
Counterproductive in what sense? Provided that TSX does provide performance benefits and universally available, I don't see why it would be unused in the software.
Also note that transistor budget grows with each node, so what is presumed to be too complex/expensive today might not be tomorrow.
> Mark Roulo (nothanks.delete@this.xxx.com) on April 9, 2021 12:54 pm wrote:
> > My employer runs a lot of hand optimized numeric code. We started with AltiVec
> > on PowerPC, then moved to SSE on x86 and are now using AVX2 on x86.
> >
> > One reason we HAVE NOT moved to Avx-512 is that the developers like being able to run code on their
> > desktop/laptop machines. Until recently, Avx-512 was pretty much only available on server parts.
> >
> > Making it difficult (or impossible) to run code on developer
> > machines discourages adoption of new capabilities.
> >
> > Note that a SLOW Avx-512 implementation on the laptop/desktop chips would have
> > been okay. The developers (other than the performance optimization) team don't
> > need the code to run QUICKLY on their machines -- they just need it to run.
>
> I get you, but it's not the same.
>
> In the case of AVX-512, it's going to see some use for average Joe. This isn't about development.
> This is about the point that for TSX to be fast enough to be worth using, it's going to require
> complexity that is going to be counterproductive to desktop/mobile. It's going to sit cold.
Counterproductive in what sense? Provided that TSX does provide performance benefits and universally available, I don't see why it would be unused in the software.
Also note that transistor budget grows with each node, so what is presumed to be too complex/expensive today might not be tomorrow.