AMD r9 Fury X

[muziqaz]

core 17 was the sweet spot for AMD, but then maxwell came in with nvidians actively optimising for the fah, so nvidia is weapon of choice in folding those cores. But AMD is not bad at them(compared to other cores).
core 18 - oh boy. openMM/opencl compilation went south for AMD. We don't know where exactly, but some insights of some beta testers devs I think know the answer to get AMD on track on this core. As it is right now, AMD is losing 2x-3x performance compared to core 17 on AMD, and compared to nvidia core 17/18, well there is no contest. Here is comparison: 7970 radeon used to kick any cpu around with core 17. But with core 18 7970 is getting similar TDP than my haswell E i7 [email protected]. 7970 is getting 3x less PPD than it used to get on core 17. So on this core nvidia hands down.
core 21 used to similar to core 18, until we figured out somewhat the issue causing performance degradation on AMD cards. So with latest core compile we are seeing AMD back in the fight, BUT without active AMD involvement to optimise for the fah, we are still behind nvidia. 466k PPD on fury x is very decent (like around 1.5x decent than 290x), BUT 980ti and titan x are getting above 500K, closer to 600k So AMD is still behind. best thing 7970 received around 3x boost from fixed core 21, which is amazing, and my haswell is again left in the dust.
As said before we moved some stones around down in AMD camp, and have some responses from certain people, so... we will see

For the record: fury x on core 18 WU is 156K- 180K ppd. I do not have HFM, so I cannot catch the benchmark data to share with you guys. BUt whenever I catch new WU I post the results. Since I finished all the fury torture sessions yesterday, this week will be folding ONLY. So hopefully will get better spread on performance.
Also keep in mind that currently in quite a few opencl benchmarks fury x is underutilised. So I expect AMD to optimise a bit more and we should see some better opencl numbers later on this summer. Cards are very inconsistent to be polite to AMD, but there are some rumours they are preparing Omega like drivers later this month/summer.

[muziqaz]

Edit: core 17 p13001 - 395k ppd

[toTOW]

Also, performance is usually better on Linux than on Windows

[7im]

You are saying GPU folding on Linux performs better than on Windows?

[billford]

You are saying GPU folding on Linux performs better than on Windows?

It can be.

I've got one machine with a GTX 780Ti which I can dual-boot into Linux or Vista- with the same graphics clock the TPF is about 10% shorter with Linux.

I also noticed that with Vista I could drop the memory clock from 7000MHz to 5000MHz without any noticeable effect, the same reduction with Linux increased the TPF by around 3-4%.

This was about a year ago with the then latest drivers- whether or not it would apply to other cards, later versions of Windows or the latest drivers I have no experience.

[toTOW]

You are saying GPU folding on Linux performs better than on Windows?

It does, at least on NV GPUs ...

[7im]

By how much? Anyone got hard numbers?

[monkeyclaw]

Thank you muziqaz for the breakdown of where Nvidia and AMD are both in terms of folding efficiency and output. It's very much appreciated, and I also hold out a lot of hope for some of AMD's people coming on board again and helping to optimize along with the F@H team. I feel like if those within the AMD camp do finally start paying attention again, we can see some huge performance gains.

Thanks for all the updated folding numbers too, keep em comin'. I'm probably going to end up getting the R9 Fury X, if not because it's a great card, then because it's more than comparable with Nvidia's offerings at the price range, and AMD I feel needs more support to stay in the fight with Nvidia... competition is good for keeping prices down, and AMD is definitely the underdog right now!

Thanks again to all posting PPDs and info about the Fury X!

[bruce]

I feel like if those within the AMD camp do finally start paying attention again, we can see some huge performance gains.

A lot depends on what you define as "huge."

[Sven]

New PPD Number from Project 10468 with 380'000 read with HFMdotNET

[monkeyclaw]

I feel like if those within the AMD camp do finally start paying attention again, we can see some huge performance gains.

A lot depends on what you define as "huge."

Huge to me would be simply on par with Nvidia in terms of teraflops -> PPD. Or, maybe even a little ahead

[kiore]

New PPD Number from Project 10468 with 380'000 read with HFMdotNET

I think there is a lot of variation of PPD between work units with these very fast modern cards, so exact comparisons difficult. However the numbers quoted here seem similar to what I get with my GTX 980ti (on win7), averaging around 400k. On the edges I see a 10486 recorded on HFM at 470k up to a 9110 at 515k, many other units are in the low 300ks like a 9128 at 327k.
Looks like the r9 Fury X pretty competitive at present for FaH.

[muziqaz]

fury x can be only called competitive once it is kicking titan x numbers around, and only then it has so much potential. price wise yes.

[monkeyclaw]

fury x can be only called competitive once it is kicking titan x numbers around, and only then it has so much potential. price wise yes.

I don't know man. Last I checked, the Titan X was still a $1000+ card, so it's not fair to say Fury X has to compete with that to be competitive

But hey muziqz, those guys in AMD that gave responses to F@H about working with them again... do you know more details about what AMD said, and how long ago they said it? I've pretty much decided I'm going for an AMD card when they finally come in stock for more than 5 minutes, but I am excited now about a possible optimization for F@H from AMDs end =)

and kiore, you are really right: there can be a LOT of variation among different projects: even with just core17 units on a kinda older/mid grade AMD card (R9 280x), I see variation. That is really interesting and reassuring though that the 980ti isn't blowing the Fury X away right now, very good to hear

Still following this thread and looking for my own Fury X to unleash on the Stanford team Keep the numbers and updates coming people! I love seeing new info here

[muziqaz]

Last blip of activity in the emails was on 27th of June, and the word was that they are still trying to wake up the guys who are kinda supposed to be involved with F@H. Unless they moved to private emailing and contacting between each other, I'm afraid there was nothing ever since.
Few things to note, before we go negative towards AMD:
AMD is in really bad financial state, and currently are trying to survive, so it is understandable that they just cannot afford manpower to even check out folding. I for sure have no clue what is going on inside AMD, and what their policies are, so I try not to judge.
The person who initially responded to my requests was John Bridgman. He is AMD Linux driver guy I think. The fact that he has nothing to do directly with what we are trying to do here, makes me give him a lot of respect, as he was very enthusiastic and willing to check things out and turn some stones inside of AMD. For that I tip my hat to him, and say thank you for even trying to do something and responding to us.
I try not to antagonize people in big companies, unless they complete dicks (which is not the case here at all), but I will try somehow subtly start the conversation again