Intel's Clear Linux Helping AMD EPYC Genoa Hit New Performance Heights

[pipe13]

Has ifort added compatibility against any of the common Clang and GCC flags? At least with DPC++ early on was going to benchmark that, but the various compiler flag differences and the time involved for ensuring consistency made it not worthwhile.

That I couldn't tell you, as your "common" Clang and GCC flags are likely to be quite obscure compared to my own. All I've used at present is
FFLAGS = -O2 -fPIC -mcmodel=large -u
so quite basic. (mcmodel=large allows code and data segments larger than 2GB.) One would have to dig through the ifort man page to see if it has exactly what you want, and it seems you've already done that. Thanks for the effort!

[mdedetrich]

It was an oversight; we set that a couple of years ago and more or less forgot about it since we never hit that limit... the phoronix article is the first complaint we got basically.
I just bumped it way up to not have to think about it for another couple of years after this ;-)

I thought it was an oversight, thanks for making the time to create a response!.

[stesmi]

While 512 is absolutely enough for even Zen4 Bergamo (up to 128 cores / 256 threads) in a dual socket configuration, that's already hitting the max there. Granted, going OVER 512 threads might take a while, but...

// Stefan

[arjan_intel]

While 512 is absolutely enough for even Zen4 Bergamo (up to 128 cores / 256 threads) in a dual socket configuration, that's already hitting the max there. Granted, going OVER 512 threads might take a while, but...

// Stefan

I'm sure it'll go over 512 at some point in the future ;-)
512 I felt safe with as a quick change... much higher and I'd like to figure out how to test first.

I would rather boot with a few less cpu cores, than crash and burn during boot and just die.
In the former case you can at least software update to a next build that has a higher setting... in the later case you're pretty dead in the water

[qarium]

And yet there are still some developers questioning the usefulness of system tuning and performance optimizations in general. They keep repeating the "O2 is enough for everybody" and "AVX2/AVX512 is a waste of silicon" mantra. I think this review proves it again that there are major gains to be had in certain workloads, and to be frank, I'd expect this performance work to be done by upstream developers and other mainstream distros, too, to provide a better out-of-the-box experience for their users.
Good job, Intel.

can you tell me why is AMD successfull with AVX512 and intel is not ? right now AVX512 is a waste of silicon if you buy an intel cpu...

and i think it is microsoft who owns ubuntu wo is sapotaging the linux performance.

it looks like compared to ubuntu Intel's Clear Linux is a very good product.

Intel could change their mind at any time and license the AMD-AVX512 double pump(not the pentium4 fouble pump) solution... intel could also change their mind and just license RDNA3...

amd with apple M1/M2 apple would benefit a lot by license RDNA3...

[stesmi]

I'm sure it'll go over 512 at some point in the future ;-)
512 I felt safe with as a quick change... much higher and I'd like to figure out how to test first.

I would rather boot with a few less cpu cores, than crash and burn during boot and just die.
In the former case you can at least software update to a next build that has a higher setting... in the later case you're pretty dead in the water

At this point, yeah! I didn't mean any disrespect, btw. I totally get why it was lifted the way it was, and I'm happy you went to 512 and not just to what was needed - 384.

// Stefan