Zen4's AVX512 Teardown

Huge deep dive into AMD’s new AVX512 instructions. He’s been under NDA for 4+ weeks, just released. His conclusion:

Overall, AMD’s AVX512 implementation beat my expectations. I was expecting something similar to Zen1’s “double-pumping” of AVX with half the register file and cross-lane instructions being super slow. But this is not the case on Zen4. The lack of power or thermal issues combined with stellar shuffle support makes it completely worthwhile to use from a developer standpoint. If your code can vectorize without excessive wasted computation, then go all the way to 512-bit. AMD not only made this worthwhile, but incentivizes it with the power savings. And if in the future AMD decides to widen things up, you may get a 2x speedup for free.

EDIT: more comments about some of the added instructions here

5 Likes

I am STRONGLY considering abandoning TMF forever. Admittedly I just finished dialysis, but the lack of a next post button and no obvious way for me to find it manually means User Idiots who have no interest in satisfying users with decades of experience here are now driving me away.

3 Likes

You probably need to think of the new boards as a game. You hunt around and find things and every so often get rewarded.

Spoiler: you got a reward!

Mike

2 Likes

The fact that they couldn’t even retain my previously read status making all posts look like I hadn’t seen them shows they don’t care about users. They dumped all PC users to make it more convenient for mobile (which I don’t think it is) shows a lack of focus groups. How hard would it be under the more option to have a Read New in Thread icon (you get the text on mouseover) and a Read Next Thread icon? There is plenty of room for two more icons, and that would add functionality the old boards lacked.

As for rewards, that’s so last century and so unFoolish. The ratio of smart to stupid people here is about the highest I’ve ever seen, and I’ve learned so much from so many of you that I am eternally grateful. Unless the rewards are meaningful and clearly explained, they aren’t going to motivate high intelligence people. For example, I buy an annual boost when I play Heroes of the Storm. It gains me nothing, but it sends Blizzard $20 for continuing to provide a free game. I pay for upscale membership on several websites (Board Game Arena comes to mind) which give me more information about whether a game REALLY benefits the skilled player (I look at the average dice rolling option and no longer like Stone Age nearly as much).

TMF is trying to attract a larger slice of the bell shaped curve. That will decrease their signal to noise ratio, but increase their pool of eyeballs and ad revenue. Instead, why don’t they add a Board User option where we pay $20-$50 a year and get access to some features such as, say, a New In Thread and Next Thread button. What features? Maybe make some of Fool’s own boards only accessible, maybe posts written by Fool Analysts. Those are all of the top of my head.

Even Facebook, after their most recent major UI change, retained an option to keep the old (non-mobile) interface. Couldn’t the smart people at TMF have learned from FB’s mistakes? Yes, we have many young users, and some of our older users have died off or just stopped posting. I don’t mind a recruiting effort. But at any given time, people here have enough spare money to do some investing, which skews the target audience for TMF older, better educated, and no more student loans. With those things tend to come weaker eyes, worse hand eye coordination, don’t trust the innate insecurity of cell phones, prefer larger ergonomic keyboards and antivirus capabilities of PCs.

Game design 101: know your target audience. That’s a first lecture concept (I’ve taught the course). Fool forgot or didn’t think about its target audience.

Next week I’ll tell everybody how to fix the world. That will take several months just to lay out the plans and get banned for politics on this board.

3 Likes

RP if you, or anyone else here, needs help, I’ll be glad to assist. By changing a few settings, it’s possible to get a similar experience to what we had before. It’s not an exact match, but close enough, IMO.

3 Likes

I have a stupid workaround that requires reloading the page to go back. But I am not happy.

Would love to hear your how-to on this new platform.

So far I have figured out that if I add “advanced-micro-devices” to my Tags I can click on that get all the posts on AMD. Then again there is now an “AMD” tag as well so I have to check both tags to make sure I don’t miss anything. It also requires that all posts are properly tagged.

As far as I can tell the new platform just dumps every post from all users into a slurry and then you hunt for things using Categories which are so broad that you get a large feed of crap or narrow the search with tags.

What am I missing because the new platform has effectively killed our AMD group. Would love to figure out how to keep it going.

This is a good primer: New Discussion Boards Help Guide

It is true that you can see “every post from all users in a slurry” if you want to, which is I guess an increase in functionality from before, though one I gather you are not interested in. I think you will find that the “advanced-micro-devices” tag is a 1-for-1 replacement for the AMD board you knew and loved. Posts from the old board were migrated over with that tag applied, and any posters wishing to participate as before should just post with that tag. If you feel this is leaving something out, could you specify what that is?

Regarding these particular tags, “AMD” is an alias for “advanced-micro-devices”. Whether you go to https://discussion.fool.com/tag/advanced-micro-devices or https://discussion.fool.com/tag/amd, you should see exactly the same thing. This is exactly so that people who think by ticker symbols and people who think by company name will not miss each other. This should be the case for any company tag.

1 Like

Why do you say the old AMD board has been essentially killed? The one that we’re posting on now IS that old board, resurrected on this new platform. Unless you mean being able to see the past posting history? I believe that’s here too, but I’m not sure how to see it!

only posts from this year were imported. Older posts are still visible on the old boards, but not for long.

1 Like

Ah, yes. I read that somewhere before, perhaps from you. And I now see that to see up-to-year-old posts, I merely have to keep scrolling down after going to Topics tagged advanced-micro-devices.

Or, rather than scrolling down, click on the Activity column heading until the oldest posts are at the top. That assumes you are in the Latest view.

Year, that quickly gets you to the oldest posts. Now, how do you quickly get to the full Topics tagged advanced-micro-devices view that shows all posts regardless of criteria? (other than just editing the url that is). I.e., if I still have some unread posts, clicking on the advanced-micro-devices tag defaults to only showing the unread posts via https://discussion.fool.com/tag/advanced-micro-devices/unread.

The controls at the top can toggle between the various sorts. For the full board, you want “Latest”.

image

Thanks. So Topics tagged advanced-micro-devices and Topics tagged advanced-micro-devices appear to be equivalent.

Sorry I haven’t posted for a bit. When getting used to/set up for the new boards, I happened upon this post. I still haven’t finished digesting it, and it is probably going to mean lots of work adding a Zen 4 code path to some computational libraries. I’ve gotten off easy with AVX2. For many of the 256-bit instructions–at least for how I would use them–sticking with the original 128-bit versions either functions the same or the 256-bit version adds a bit of latency. But this says that the AMD AVX512 should be used–assuming you have a Ryzen 7000 or Genoa chip. I feel like I just finished installing my R9 5900X, and upgrading will be a lot of work. Also, I need to keep a Zen 3 system for comparison. I guess I could say Merry Christmas to my son and have him run benchmarks from time to time.

Also Der Bauer has been at delidding again. He says he gets on the order of 20 degrees C drop on the 7950: https://overclock3d.net/news/cpu_mainboard/der8auer_achieves_unreal_temperature_improvement_with_direct-die_cooling_on_ryzen_7000_cpus/1

I’m not going to tell you to delid your CPU, and I have no plans to do it. But AMD could provide new Zen 4 SKUs with a space around the pads on the bottom of the chip instead of the thick heat spreader. Easy way to get improved performance without pain. Or maybe just offer chips with a solid silver heat spreader, or even a vapor chamber. :wink:

2 Likes

I’ve never understood why relatively simple measures like this aren’t put into practice given the success people seem to have in the aftermarket. Compared to the other measures chipmakers spend billions on to squeeze performance out of the silicon, they seem like low hanging fruit…

1 Like

Compared to the other measures chipmakers spend billions on to squeeze performance out of the silicon, they seem like low hanging fruit…

Want to get rich? Design a heatspreader which uses carbon nanotubes embedded in copper, silver, or perhaps even graphite to increase the heat conductivity of the heatspreader. Patent it, and watch the royalties roll in. Better yet, patent the manufacturing process as well, and license that, or go into the manufacturing business.

It would probably take a small lab and a few months of experimentation to find the best design. You may want to experiment with alloys with a higher melting point than the solders used in CPUs but much lower than pure copper. Start with lead and mix some tin in to get an alloy that won’t harm the nanotubes. I’m not sure whether it would be worth sorting for only conducting nanotubes.

1 Like

Looks like it has already been done 15 years ago, in various ways:

1 Like

Actually no. I could go through the patents individually, but they all have to do with growing nanotube structures in situ. (To replace the bumps, to replace the TIM,* and on top of a vapor chamber.) I am suggesting constructing a heat spreader material that can be cut to fit. Think of a cookie cutter process. Mix the CNTs with the metal, pour on a flat surface, roll to required thickness, then cut out individual heat spreaders. (Or you could cut into strips then cut the strips to length.)

In any case, creating the heat spreader will have no effect on the yield of the chips. And the cost of the heat spreader might be a few cents less than the current copper HSs.