Upcoming AMD gpus and big navi

[MeeLee]

From tflops count of the upcoming AMD gpus, 2 out of 3 have already been verified by outside sources.

The codename Navi 22 GPUs are expected to have around 12.8Tflops, which lands them in RTX 2080 Super territory.

Navi 21 is expected to run 22.5 Tflops, or somewhere between an RTX 3070 and 3080.
The leak I read, did describe that tflops and performance are harder to calculate on Nvidia GPUs, as they can do both INT and FLOAT operations, and switch rapidly between them.
If the software isn't optimized for that dual shader calculations, there's a chance that Navi 21 might be faster than a 3080, simply because the 3080 might not reach the promised 30Tflops.

No concrete info on TDP, but rumors are that Navi 21 could come close to 400W TDP!
And that this is thanks to the high shader count, as well as high gpu core count (2.2Ghz).
I'm sure a lot of that will be marketing, and sustained boost will be like Nvidia, below 2Ghz. But even if, one can save a lot of power/heat by just underclocking the gpu by 100-200Mhz (=~10%) on the core.
Those last 10% of performance are responsible for the highest power draw...

[foldy]

Biggest problem is high shader count of AMD Navi 21 with 5120 shaders. FAH cannot use so much shaders especially on AMD GPUs with FAH work units having less than 100k atoms count.

[psaam0001]

Looks like I'd better budget for a 1500+ watt power supply, and a licensed electrician when I am able to build my next "COVID Buster". Especially if I plan to use more than 2 of these new cards on the same 64+ core AMD Threadripper [SIC] system (cabinet space permitting).

Paul

[JimF]

It looks like some of the new CPU Core A7 work units in the 174xx series have pretty big atom counts (165 k).
https://apps.foldingathome.org/psummary

It may be a while before a new card is really necessary.

[psaam0001]

But AMD needs to make their next set of GPU driver packages for the major Linux distributions easier to install (hint: better documentation). Till then, it's NVidia cards for all of my future Linux builds

At least w/NVidia, I know what I need, and can find help to install it.

Paul

[MeeLee]

Nvidia will be better regardless, even if their performance is slightly lower; their efficiency is expected to be heaps above AMD, as they're at least 1 generation ahead of AMD.
Even with the 'older' RTX 2000 10nm process, AMD can't bounce up to their efficiency with 7nm designs.
If we look at the RX 5700 XT, at ~2500 shaders, and 225W at 7nm, it still manages to underperform similar Nvidia's 10nm GPUs.
Which leads me to believe that AMD will mainly jump on the software side, to, like Nvidia, will make sure data doesn't hop around too much inside the GPU (to increase efficiency), starting with their Big Navi GPUs.


Bigger atom counts will be expected in the near future, or processing two or more projects per GPU.
Big Navi isn't expected for at least another few months. We'd be lucky to have one in our hands in 2020. But I think they're gunning for the Christmas season.
Even then, there's a possibility many will just be able to get their hands on them in 2021.

I'm expecting people running Linux, to be the first to have issues with the GPUs, mainly software issues.
AMD has a lot of catching up to do in the GPU sector.
I felt that Nvidia's RTX 3000 was a pushed event. The technology wasn't thoroughly tested, and they had no competitors on the market.
However, their newer tech would mean lower manufacturing cost (GDDR6X), and I guess it made sense to prematurely discontinue the 2000 series GPUs.
They easily could have just done like Intel, and kept on selling RTX 2000 GPUs, as there were no contenders; meanwhile, invest more time in R&D.
But they didn't.
It's between a risky or a wasteful choice they had to make, and the risky choice backfired on them...
Can't blame them. Sometimes risky choices reward, sometimes they backfire.
From what I gather, the RTX 3000 woes is mainly because of lower quality capacitors on the back of the GPU, of which there's no easy fix (one could solder additional capacitors manually on the GPU, if one had a fine hand, and knew what needed to be done).
I'm presuming that the caps in the back were for 2000 hardware, and the newer design has some sort of additional computing to do, to send 4 signals to the GDDR6 memory modules, rather than 2 on any older GPUs released to this date.
The importance of 2 bit memory is vast! If this translates to 2bit GPU cores in the near future, means that GPUs will be able to run much more efficiently (as now there are 4 possible states per clock, instead of only 2).

The hardware problem, is something AMD has learned from, and probably will do more testing on their GPUs, and make sure third party manufacturers will not sell their GPUs with similar flaws, and potentially may push back the release date.
I don't think we'll see these hardware problems on AMD at the release, but there might be potential software issues.

Looks like I'd better budget for a 1500+ watt power supply, and a licensed electrician when I am able to build my next "COVID Buster". Especially if I plan to use more than 2 of these new cards on the same 64+ core AMD Threadripper [SIC] system (cabinet space permitting).

Paul

You can easily run 3 Big Navi GPUs on one socket!
They're 400W max, but you could run them lower.
In fact, just running FAH, they probably will run lower.
At 375W per GPU, x3 = 1125W. Add to that a cheap dual core Celeron CPU and motherboard, and you'll be running at 1350W. Any 1600W PSU can easily run it, on a 15A breaker line.
Personally, I'd not go over 2x Big Navi, mainly to get x8 speed slots, and secondly because single phase (120V) 1600W PSUs are somewhat exotic and expensive.

[psaam0001]

2 RTX 3090's would do... My GT 1030 can fill that third slot when my dual core Core2 CPU system has folded its last WU.

Paul

[Frontiers]

There is some leak from macOS 11 beta :
https://www.reddit.com/r/Amd/comments/j ... ?context=3
Looks like 80 CUs Navi21 will have 2 flavours, with 2050 and 2200 MHz boost clocks.
Last clock is out of reach for GA102, first maybe with good custom watercooling.

[bruce]

So both will be out of reach for production hardware??? That's not good.

Or are they going to sell water cooling?

[MeeLee]

There is some leak from macOS 11 beta :
https://www.reddit.com/r/Amd/comments/j ... ?context=3
Looks like 80 CUs Navi21 will have 2 flavours, with 2050 and 2200 MHz boost clocks.
Last clock is out of reach for GA102, first maybe with good custom watercooling.

I don't believe it, unless I see it.
AMD has somewhat lost my trust, with their Ryzen 3000 CPU line.
They supposedly also ran 4,7Ghz boost, and 4,5Ghz on all cores, something I couldn't even begin to get close to!
Highest I ever saw was 4,125 Ghz (peak) while most of the time that same CPU runs 100% workloads at 3,85Ghz (3,9Ghz on a good day).

I expect the same for their GPUs (2,2Ghz will likely be 2Ghz, and 2Ghz will likely translate to 1,8Ghz, which is what we have with Nvidia now).

[muziqaz]

There is some leak from macOS 11 beta :
https://www.reddit.com/r/Amd/comments/j ... ?context=3
Looks like 80 CUs Navi21 will have 2 flavours, with 2050 and 2200 MHz boost clocks.
Last clock is out of reach for GA102, first maybe with good custom watercooling.

I don't believe it, unless I see it.
AMD has somewhat lost my trust, with their Ryzen 3000 CPU line.
They supposedly also ran 4,7Ghz boost, and 4,5Ghz on all cores, something I couldn't even begin to get close to!
Highest I ever saw was 4,125 Ghz (peak) while most of the time that same CPU runs 100% workloads at 3,85Ghz (3,9Ghz on a good day).

I expect the same for their GPUs (2,2Ghz will likely be 2Ghz, and 2Ghz will likely translate to 1,8Ghz, which is what we have with Nvidia now).

80CU has been circulating since the beginning. Clocks - since Xbox/ps5 announcements. 50% power improvement since the beginning as well mentioned by AMD themselves.
Big Navi if all goes to plan should be very good in gaming, much better showing as to what sh*t show nVidia done up till now.
But we know AMD has an ability to turn gold into complete crap when marketing it, so, yeah.

When it comes to folding, 40 CU Navi is decent at folding, much more efficient than any GCN based AMD card, but when we double that CPU count we end up with starving CUs again, I believe, so again we will need huge Atom counts in order to utilise all that power, so we will be at square one again, as we are with GCN (Radeon 7/Vega64). Unless AMD decides to bring a person or two into the circles of fah developing, that it sthe future of AMD in folding I'm afraid.
nVidia completely ran away with CUDA, and only heavy AMD driver team involvement would help them bridge the gap.

Oh, and Navi21 at 400W - complete nonsense. We have nVidia blowing their circuits. One is enough

[NormalDiffusion]

When it comes to folding, 40 CU Navi is decent at folding, much more efficient than any GCN based AMD card, but when we double that CPU count we end up with starving CUs again, I believe, so again we will need huge Atom counts in order to utilise all that power, so we will be at square one again, as we are with GCN (Radeon 7/Vega64). Unless AMD decides to bring a person or two into the circles of fah developing, that it sthe future of AMD in folding I'm afraid.
nVidia completely ran away with CUDA, and only heavy AMD driver team involvement would help them bridge the gap.

Yes, it's gonna be problematic for AMD to catch up now with the CUDA improvement... I just got a 1080Ti and I'm crying when looking at what she's doing in comparison to my R VII's and Vega FE...

[bruce]

nVidia has put a lot of development money into giving themselves a competitive advantage with CUDA compared to the performance of the static OpenCL code. I doubt AMD has invested that much into their drivers ... and as you say, adding CUs that are only beneficial on proteins that have lots of atoms has limited value for either company.

[muziqaz]

Bruce, regardless how much nVidia invested in CUDA, their efforts to optimise for FAH and Adam's involvement with the project has been super example. If AMD had someone constantly working with us, we would see similar performance from AMD GPUs, I'm sure of it. As it is right now AMD are folding on rather generic OpenCL code, which was only tuned by FAH devs, and seen no major involvement from AMD driver/software team.
And seeing how FAH received so much attention in recent months, nVidia now is reaping the benefits of their involvement with the project, and good for them

[NormalDiffusion]

If AMD had someone constantly working with us, we would see similar performance from AMD GPUs, I'm sure of it.

Maybe it's time to contact them and directly ask if they don't want to help a little bit more... They are still listed as sponsor, and I don't think it would hurt to have the fah code optimized (OK, before launch of the new GPUs probably not possible).
I have to admit, I'm regretting my choice now (3x r vii + 1x vega 64 fe lc) when I see what an old 1080ti or titan xp are doing right now . I'll closely follow the development in the next weeks and months, and buy my new gpu(s) regarding the performance in fah. And I don't think I'll be alone on this track!