mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   y-cruncher (https://www.mersenneforum.org/forumdisplay.php?f=159)
-   -   AVX512 and Zen4 pre-release speculations (https://www.mersenneforum.org/showthread.php?t=28032)

Xyzzy 2022-08-28 18:43

AVX512 and Zen4 pre-release speculations
 
[URL]https://www.techpowerup.com/298194/latest-y-cruncher-version-comes-with-zen-4-and-avx512-optimization[/URL]

:mike:

Mysticial 2022-08-29 21:28

I'll have a lot more to say once the embargo lifts. But it looks like some of the big names also have the chip. So the tear-down of the chip which I've prepared probably won't be the only one - let alone the best. Especially since I'm mostly a SIMD person with less insight into the rest of the chip.


Lots of juicy stuff to come whenever I'm allowed to post. :razz:

mackerel 2022-09-03 13:45

Looking forward to it. A problem with the more well known reviewers is they try to cover everything, and FP perf doesn't get much depth to it, or is tested in ways hard to relate to our interests. For the audience here more info in that area would be very interesting.

LordJulius 2022-09-04 04:24

RE: Zen4 & AVX-512
 
[url]https://hothardware.com/news/amd-ryzen-7000-zen-4-launch-speeds-specs-all-you-need-to-know-and-an-rdna-3-surprise[/url]


"Some details and performance expectations were also disclosed regarding Zen 4’s AVX-512 implementation. In Zen 4, AVX-512 is implemented using double-pumped 256-bit data chunks. This design decision was reportedly made to avoid large frequency fluctuations when executing AVX-512 workloads. In terms of performance, AMD is claiming a 1.3x improvement in FP32 inferencing workloads versus Zen 3, and up to a 2.5X improvements for Int8."

Mysticial 2022-09-04 07:53

[QUOTE=mackerel;612593]Looking forward to it. A problem with the more well known reviewers is they try to cover everything, and FP perf doesn't get much depth to it, or is tested in ways hard to relate to our interests. For the audience here more info in that area would be very interesting.[/QUOTE]

The other problem with launch day reviews is that they fail to capture the new product with optimizations for it. Simply because the developers of the benchmark/game have not had the opportunity to do it yet.

In a way, all new products are inherently disadvantaged from start. Only months later would things actually improve - assuming the benchmark/game is still in development. That's besides the point for early adopters wanting the performance on day one, but it doesn't tell the whole story.


Zen4 feels a bit different this time as there's an unknown number of (individual) software devs who have gotten the chip very early - well ahead of the usual press who do the hardware reviews.

I'm actually surprised this doesn't happen more often. Prior to a launch, the vendor sends out samples to all the still-in-development benchmarks to let them time to do pre-launch optimizations. Makes the launch day reviews look better. Heck, get some free beta-testing in the process. Slap everyone with NDAs and no one would risk their reputation to leak anything.


[QUOTE=LordJulius;612631][URL]https://hothardware.com/news/amd-ryzen-7000-zen-4-launch-speeds-specs-all-you-need-to-know-and-an-rdna-3-surprise[/URL]

"Some details and performance expectations were also disclosed regarding Zen 4’s AVX-512 implementation. In Zen 4, AVX-512 is implemented using double-pumped 256-bit data chunks. This design decision was reportedly made to avoid large frequency fluctuations when executing AVX-512 workloads. In terms of performance, AMD is claiming a 1.3x improvement in FP32 inferencing workloads versus Zen 3, and up to a 2.5X improvements for Int8."[/QUOTE]

The double pumping was entirely predictable. Nobody expected AMD to spend that much silicon on something that's hardly used yet. (along with the power implications of full 512-bit)

Besides all the new non-width related features of AVX512, there are advantages of running 512-bit on 256-bit hardware. Half the number of instructions for the same amount of work - and therefore half the front-end overhead.

In order to fully utilize Zen3's 4 FPU pipes, you need to sustain 4 instructions/cycle. This is hard to do outside of synthetics. 4 IPC is hard to sustain in general because you're pushing against the limit of the instruction decoding and dispatch. Floating-point has long latencies and you only have 16 registers. If you play games with renaming to get around the limited # of regs, you start running up against the limit of the reorder window.

Mysticial 2022-09-15 01:08

Just wondering:

Where should I post my Zen4 AVX512 breakdown when the embargo lifts?

If I post it here, it's already buried beneath a bunch of posts. If I post a new thread in the Hardware subforum, I won't be able to fix/annotate errors after the edit grace period. If I post a new thread under this section, the title will seem redundant of this one.




Side Note: I still don't have a solid date for when embargo lifts. The date that my AMD contact gave me is almost a week before the date that I'm reading on Twitter and from my various media contacts. So when launch gets closer, I'm going to ask my contact again.

VBCurtis 2022-09-15 03:05

I'd post it right here in this thread. Anyone browsing this subforum will see the new post in the thread, and anyone who reads the site via "New Posts" will also see it. The thread title catches plenty of attention from those who like this sort of thing.

Mysticial 2022-09-15 03:15

Problem is that I'll be linking externally. :lol:

VBCurtis 2022-09-15 03:16

By all means, make your own clean thread!
If there are edits you wish to make later, you can get mod attention and one of us (I volunteer, at minimum) will paste in the edits you request.

Prime95 2022-09-15 03:28

[QUOTE=Mysticial;613461]
If I post it here, it's already buried beneath a bunch of posts. If I post a new thread in the Hardware subforum, I won't be able to fix/annotate errors after the edit grace period. If I post a new thread under this section, the title will seem redundant of this one.[/QUOTE]

How about renaming this thread to "AVX-512 and Zen4 pre-release speculations",
then start a new thread "AVX-512 and Zen4 details"?

Or whatever you want to do. Those that care will find your posts!

Xyzzy 2022-09-15 11:27

[QUOTE=Mysticial;613461]If I post a new thread in the Hardware subforum, I won't be able to fix/annotate errors after the edit grace period.[/QUOTE]Are you sure about that?

:smile:


All times are UTC. The time now is 01:45.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, Jelsoft Enterprises Ltd.