AMD cards

[Sparky]

What's it looking like on AMD GPUs these days with F@H? I have a 7950 that is currently on another project but it is ending within the next couple of days. Does it still take a whole core to feed the card? How's the computer usability with F@H running at the same time? What kind of performance are we looking at?

I tried running F@H on my GTS250 on my office PC but it made the PC entirely unusable, so I couldn't have that

[Swiftdeathz]

Not positive but AMD cards generally always trailed behind Nvidia. I think there was some new core17 that was putting out ridiculous ppd, but it's still in beta.

[Andreas]

To my understanding, the objectives of the new core (17) are significant.
1) Change that the only interface to GPU's will be based on OpenCL. The OpenCL drivers of AMD are at v1.2, NVidia is at 1.1, but both seem to run.
2) With the change to OpenCL, there seems to be the intention, that GPUs will be treated like SMP machines which get bonuses for taking on large problems.
3) This and the rest is still in the air, as everything is in beta.

Andy

[Sparky]

Yeah historically nvidia has killed AMD at F@H but wasn't sure if that was still the case, or still going to be, or whatever.

It'll be interesting on the OpenCL change as AMD has traditionally supported OpenCL much better than nvidia (they want to push CUDA for everything). I wonder how that will affect things?

[Swiftdeathz]

I read reports of 7800,7900 series cards were getting 48-50k ppd on the core17 wu's which were in beta this month.

[bds71]

Not positive but AMD cards generally always trailed behind Nvidia.

Yeah historically nvidia has killed AMD at F@H

I heard differently: a review over at Anandtech has the 7990 pretty much stomping all over both the 690 AND titan (for GPU compute) - and one of the benchmarks they performed was the F@H bench....check out the review here. (click here if you just want to see the compute comparison)

actually, I want to say that since AMD started using OpenCL they became at least as good/better than nVidia at compute (since nVidia was pushing their own CUDA platform, they didn't put much work into OpenCL) - at least for those type programs that didn't use CUDA...

[Sparky]

Well that's more recent. Back a few years ago AMD/ATI cards couldn't hold a candle to nvidia's cards. But if things have changed then SWEET

[Gimmpy224]

I heard differently: a review over at Anandtech has the 7990 pretty much stomping all over both the 690 AND titan (for GPU compute) - and one of the benchmarks they performed was the F@H bench....check out the review here. (click here if you just want to see the compute comparison)

actually, I want to say that since AMD started using OpenCL they became at least as good/better than nVidia at compute (since nVidia was pushing their own CUDA platform, they didn't put much work into OpenCL) - at least for those type programs that didn't use CUDA...

Yup youre 100% correct. And so is sparky, case and point is the 5850hd i currently use. It vs the comparable nvidia card in F@H, my card gets murdered. But since most compute programs are using opencl now instead of CUDA (like bitcoin mining, f@h, etc) the ati cards are pulling away and are able to give better bang for buck.

[Sparky]

OK well I fired it up, let's see what my 7950 @ 1GHz will do. Weird that the CCC only shows 55% GPU utilization however, and only showing 3100ppd My old GTX260 would eat it for lunch!

[Gimmpy224]

OK well I fired it up, let's see what my 7950 @ 1GHz will do. Weird that the CCC only shows 55% GPU utilization however, and only showing 3100ppd My old GTX260 would eat it for lunch!

yup, my 5850 uses 33% at 2k ppd.

i have no clue why either

[Sparky]

I recall back in the old days on ATI there were some environment variables to play with. I wonder if that still holds true.

*edit* did more digging and found it to be the version of drivers and the SDK runtime. I have 12.11 on my system as that worked best for what I was doing previously, but F@H doesn't like it. Figures. Time to do a driver dance maybe, I don't know.

I forgot how frustrating this can be with how they use this SDK stuff that is so twitchy.

[Sparky]

Can't wait for the OpenCL and core 17 units to become standard. This core 16 is killing me.

Plus, ever since I started folding again I've had a number of BSODs. I've not had that issue up until now. Stopped folding today and the PC has been happy again. I was running the GPU much harder 24/7 on another project (was loading the GPU at near 100%) so I don't get what the issue is.

F@H why you hate me

[yojimbo197]

Can't wait for the OpenCL and core 17 units to become standard. This core 16 is killing me.

Plus, ever since I started folding again I've had a number of BSODs. I've not had that issue up until now. Stopped folding today and the PC has been happy again. I was running the GPU much harder 24/7 on another project (was loading the GPU at near 100%) so I don't get what the issue is.

F@H why you hate me

Just set up my i5/HD7770 rig for Folding again today. Getting about 7K/day as an estimate for the HD7770. How many cores does the CPU project use?

[Sparky]

Yeah that is totally unfair, I was estimated to get half that with my 7950. Driver related I know, but screwing around with that ended up ticking me off lol. Sigh.

[yojimbo197]

Well now the estimate for the HD7770 GPU project is down to 2200, so my card is scraping the bottom like the rest of yuse guys.

[bds71]

there's been another rash of bad WU's this week - i'm seeing CPU WUs with 900 (and change) base points. granted, with CPU WUs, we get bonus points, so the predicted points is still over 3k, and they only take a few hours to complete...but yea - crappy WUs. the GPU WUs aren't much better - my predicted ppd is in the 75-80k ppd range (vs 120-130k with decent WUs)