Upcoming AMD gpus and big navi

[bruce]

There has been management to management contact. How an organization responds, even if they're listed as a sponsor, is up to them. FAH attempts to be vendor agnostic.

Adam's involvement has not been continuous for all of FAH's 20 years.

[NormalDiffusion]

This was quite clear for me
It's still a nice move from Nvidia and I can only hope that amd will follow!

[muziqaz]

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!

That has been done repeatedly ever since COVID struck us. By me and by people within FAH. We had this OpenCL bug in the beginning of pandemic, so that was another reason to ask them constantly. They did manage to fix the bug, but that's where their contributions ended. I since asked few more times through various outlets, with no success. Compared how Nvidia is approaching this project full steam, there is no surprise AMD is left in the dust. So even if I hate nVidia as a company, I cheered when Jensen talked about FAH in RTX3000 series presentation, and that was well deserved. FAH is large market to ignore. But hey if AMD doesn't want to be more involved, that's their choice

[muziqaz]

There has been management to management contact. How an organization responds, even if they're listed as a sponsor, is up to them. FAH attempts to be vendor agnostic.

Adam's involvement has not been continuous for all of FAH's 20 years.

We cannot expect one person to be involved with FAH all the eternity. But how he got himself involved recently is very commendable, and we see the benefits now

[NormalDiffusion]

They did manage to fix the bug, but that's where their contributions ended. I since asked few more times through various outlets, with no success. Compared how Nvidia is approaching this project full steam, there is no surprise AMD is left in the dust. So even if I hate nVidia as a company, I cheered when Jensen talked about FAH in RTX3000 series presentation, and that was well deserved. FAH is large market to ignore. But hey if AMD doesn't want to be more involved, that's their choice

That's a real pity...
And I feel with you. Long time I hadn't bought a Nvidia for the same reason you're writing... But I got myself an used 1080ti, and I was astonished by the performance with the new core...
I wanted to wait for the new amd GPUs... But now, I may switch to the green team, as performance / € is deep in their side. And with the 3070 coming end of the month... Well...

[bruce]

I don't know the specifics of any of the AMD bugs that have been fixed but I do know that FAH's OpenMM team have been in direct contact with the driver bug correction teams of both AMD and nV. Sometimes that leads to a new set of drivers from AMD or nV and sometimes it leads to a change to OpenMM. It's not easy to tell if the latter leads to a bug fix in OpenMM or it leads to a work-around that OpenMM could install to bypass the problem. I do believe that our OpenMM team is a lot quicker to respond than the AMD driver team, but it does require a new revision of FAHCore_22 to link it with a new version of OpenMM. You're in a better position to decide if it was fixed by a new FAHCore_22 or a new driver release from AMD.

Deciding whether to call it an OpenMM bug-fix or a work-around would allow us to decide who to blame, but in the long-run that gains nothing as long as the problem goes away. FAH doesn't need to do that. We'll let POTUS be the commander-in-chief of blame assignments.

[muziqaz]

We had some issues with OpenCL on Species 5 at certain Atom count projects. I don't remember if AMD actually fixed it, but I do remember getting their attention, and them being involved in fixing it.

For a moment, if a new folder has AMD card, it is good that they are cards are supported to fold, but if one asks what to get specifically for folding, nVidia is clear answer.

On the more optimistic side of things, RDNA2 architecture has some interesting design decisions, and who knows, it might work OK in folding

[MeeLee]

They did manage to fix the bug, but that's where their contributions ended. I since asked few more times through various outlets, with no success. Compared how Nvidia is approaching this project full steam, there is no surprise AMD is left in the dust. So even if I hate nVidia as a company, I cheered when Jensen talked about FAH in RTX3000 series presentation, and that was well deserved. FAH is large market to ignore. But hey if AMD doesn't want to be more involved, that's their choice

That's a real pity...
And I feel with you. Long time I hadn't bought a Nvidia for the same reason you're writing... But I got myself an used 1080ti, and I was astonished by the performance with the new core...
I wanted to wait for the new amd GPUs... But now, I may switch to the green team, as performance / € is deep in their side. And with the 3070 coming end of the month... Well...

Considering that many people still aren't able to get their hands on a 3080, I'm expecting the 3070 to be at least 3 months away from now.

[MeeLee]

For a moment, if a new folder has AMD card, it is good that they are cards are supported to fold, but if one asks what to get specifically for folding, nVidia is clear answer.

Nvidia always has been the answer.
Not only driver optimized, but also power efficiency.
AMD can't beat any of this.
There are some projects on Boinc that benefit from AMD GPUs (64bit DP calculations), allowing one to run as much as 7 projects per GPU at a time (vs 3-4 on a 2080).
But for Folding, it's been Nvidia.

[muziqaz]

For a moment, if a new folder has AMD card, it is good that they are cards are supported to fold, but if one asks what to get specifically for folding, nVidia is clear answer.

Nvidia always has been the answer.
Not only driver optimized, but also power efficiency.
AMD can't beat any of this.
There are some projects on Boinc that benefit from AMD GPUs (64bit DP calculations), allowing one to run as much as 7 projects per GPU at a time (vs 3-4 on a 2080).
But for Folding, it's been Nvidia.

Before CUDA, if a folder didn't want to buy Nvidia or Nvidia cards were more expensive overall, you could make a case for AMD in folding. Now that is not the case by far

[NormalDiffusion]

Nvidia always has been the answer.
Not only driver optimized, but also power efficiency.
AMD can't beat any of this.

That's not totally true. Take the radeon vii: it was doing fine before nvidia got the cuda update, could compete with 1080Ti, Titan Xp and 2070 Super (I'm only referring to cards I have direct access to). On 13426, the rvii is doing 1.9/2 MPPD. The average chip consumption last night (~12hrs) was around 112W on a mix of 13426 and 174xx. Given, the card is undervolted (why not when I get the same performance with less electricity?), but still more efficient than a 1080Ti or Titan Xp (haven't ever checked the power consumption on the 2070S). Couple of months ago, you could get the rvii new for 570 euro (taxes included)...

[ThWuensche]

That's not totally true. Take the radeon vii: it was doing fine before nvidia got the cuda update, could compete with 1080Ti, Titan Xp and 2070 Super (I'm only referring to cards I have direct access to). On 13426, the rvii is doing 1.9/2 MPPD. The average chip consumption last night (~12hrs) was around 112W on a mix of 13426 and 174xx. Given, the card is undervolted (why not when I get the same performance with less electricity?), but still more efficient than a 1080Ti or Titan Xp (haven't ever checked the power consumption on the 2070S). Couple of months ago, you could get the rvii new for 570 euro (taxes included)...

There is an initial port of openMM to HIP, AMDs open source equivalent to CUDA. I have built it and run F@H WUs on it. That port however is not yet optimized. The guy from AMD, who made the port, first helped me to get it to build, but I did not get feedback so far on whether AMD is willing to put ressources into optimizing it. They argued they did that in support of the upcoming Frontier and El Capitan supercomputers. Their ROCm open compute framework so far supports Linux only, since that is what is run on the platforms they target. Let's see, whether there will be an answer ...

[MeeLee]

Nvidia always has been the answer.
Not only driver optimized, but also power efficiency.
AMD can't beat any of this.

That's not totally true. Take the radeon vii: it was doing fine before nvidia got the cuda update, could compete with 1080Ti, Titan Xp and 2070 Super (I'm only referring to cards I have direct access to). On 13426, the rvii is doing 1.9/2 MPPD. The average chip consumption last night (~12hrs) was around 112W on a mix of 13426 and 174xx. Given, the card is undervolted (why not when I get the same performance with less electricity?), but still more efficient than a 1080Ti or Titan Xp (haven't ever checked the power consumption on the 2070S). Couple of months ago, you could get the rvii new for 570 euro (taxes included)...

It was my understanding that AMD was at least 1/4th to 1/3rd behind on Nvidia on the power consumption numbers, despite them being on 7nm.
Many GPUs on other platforms were tested and the lowest a high end AMD GPU would go is 200W, vs 160W for a Nvidia GPU.
The 2070 can run on 125W for the Rog Strix, and 129-135W for any other models, performing within 2% of performance on stock power.

[bruce]

John has been directly involved with this -- mostly behind the scenes. In searching for a set of WUs that benchmark all aspects of all hardware, they've encountered a variety of crashes or less severe errors and in many cases, those can be brought back to the lab and reproduced. Once a bug can be reproduced, it can be fixed and like I've said above sometimes that leads to an upgraded OpenMM and sometimes to upgraded drivers. The driver updates come from nV or from AMD, of course, and when our people are communicating directly with their people, it works better.

[foldy]

AMD RX6000 GPUs are announced for Christmas 2020 with double performance than RX 5700.