Hi guys/girls,
It's been a long while since we've had any news regarding AMD GPU crunching. Surely since the HD 7000 series release a year ago there have been improvements to AMD's OpenCL runtime?

Hi guys/girls,
It's been a long while since we've had any news regarding AMD GPU crunching. Surely since the HD 7000 series release a year ago there have been improvements to AMD's OpenCL runtime?

Between drivers 12.10 and 13.1 there have been 2 big speed increases in AMDs Open_CL. The first bump was about 20% and the second bump was about 10%. Using anything later than 12.11 beta 8 now makes the equivalent (to NVIDIA) AMD cards the fastest at POEM.

Using anything later than 12.11 beta 8 now makes the equivalent (to NVIDIA) AMD cards the fastest at POEM.

I've been "nVidia fanboi" for a long time but price and income realities are forcing me to look again at AMD. Just curious, how are you assessing equivalency in the above statement, equivalent price? equivalent numbers of "cores" or whatever the proper term is? (Of all the sub-topics in BOINC, GPU is the one I know least about, trying to catch up tho').

I can't speak for Beyond, but I think he determines equivalency by looking at the top single GPU AMD and Nvidia cards, which are the 7970 and the 680, and saying those are "equivalent" and then the 670 and the 7950 are, and so on. Before AMD dropped their prices, I'm pretty sure the 7970 and 680 were very close in price and close in performance, so they could be considered equivalent. However, I've heard AMD GPU's can overclock more and they are cheaper in price, so some can argue that GPUs like the 7970 and 680 aren't equivalent.

Yeah, I'd base it roughly on price as well. If one consumes significantly more power this might also need to be factored in (for crunching), but in the current generation both are rather efficient.

MrS

---

Scanning for our furry friends since Jan 2002

Using anything later than 12.11 beta 8 now makes the equivalent (to NVIDIA) AMD cards the fastest at POEM.

I've been "nVidia fanboi" for a long time but price and income realities are forcing me to look again at AMD. Just curious, how are you assessing equivalency in the above statement, equivalent price? equivalent numbers of "cores" or whatever the proper term is? (Of all the sub-topics in BOINC, GPU is the one I know least about, trying to catch up tho').

Pretty much what Dylan and ETA said. Also take a look at the top computers at POEM, dominated by AMD GPUs. There's quite a number of other projects where that's also true.

Thanks for the info, all of you.

Mmmmm, I see... the 60 top POEM hosts are all AMD and all Windows too. I see only 1 using AMD on Linux. IIRC, the AMD drivers for Linux were not that great a while back. It has to work on Linux, not work perfect but decent. Has that situation improved or was it never as bad as I thought it was? Troublesome installs are troublesome but never a show stopper for me, if it's good it's worth fighting for.

Got the dough ready for a GTX 690... real dough not plastic. Looks like I can get 2 AMD 7970 for roughly the same money. Is that the top of the line AMD... 7970?

There's an official overclocked 7970 called the 7970 GHz Edition also offered by AMD, and a 7990 card made by other companies that is like the 690, where it has two dual (7970) GPUs in it.


Check for yourself.


http://www.amd.com/us/products/desktop/graphics/Pages/desktop-graphics.aspx


P.S. The link doesn't open a new tab.

IMO the dual GPU cards (both NVIDIA & AMD) have had a troublesome history. I'd recommend 2 7970s over a dual GPU model. The 7970 has also come down in price a lot.

What sort of troubles?

I have a 7990 and 1KW PSU in my shopping cart at Newegg and would have attempted a checkout but I know I am $5 shy in my bank account. Tomorrow I'll deposit cash to make up the difference. Maybe that was a fortuitous delay.

I've been planning a purchase for months and have been torn between various high end CPU combinations along with GPU. Recently I decided my strategy would be cheap low end CPUs with top o' the line GPUs since GPUs have such incredible compute power. The cheaper mobos, for example the one I already own and plan to put a GPU into, have but 1 PCI-E x16 slot so a dual GPU card is the only way to put 2 GPUs onto such a mobo.

The next mobo I buy will likely then have 2 x16 slots.

If I could put a GPU on a Rasp-Pi I would do it, assuming there was the required app and other pre-requisites. Just enough CPU to drive a whopping GPU... that's the goal. I think there might be a niche market there, small but somewhat profitable, for anyone who could engineer a $100 "GPU driver" that would have 0 frills and just enough to drive one or two GPUs. I'm talking NFS storage, IPX boot and not much more. The target market would be BOINCers, the Folding@home crowd and distributed.net crowd. There may be others. Send me schematics, I might be able to prototype it in my PCB shop.

> What sort of troubles?

VRMs overheating and driver problems that don't plague single GPUs. Maybe they've ironed things out in the new generation but doubt it. These problems are true for both AMD & NVIDIA.

For $800 AR you can get 2 of the 7970 1000MHz XFX cards with the double lifetime warranty:

http://www.newegg.com/Product/Product.aspx?Item=N82E16814150586

The 7990 will cost you $100 more and only runs at 900/925MHz with very little OC headroom (the 7970 should OC far better). Just my opinion...

Maybe a little bit about brands from just one persons experience. I've been running an average of 16+ GPUs for the last few years. I've had 4 ASUS cards, of those 1 had a fan failure and 2 5850 cards had complete failures. They were sent in for RMA 2 months ago and I still haven't received replacements. The 4th ASUS I sold. Had a couple of EVGA cards, they're still OK. One Diamond, OK. One Sapphire, OK. Three MSI, all OK. Three Galaxy, all OK. One HIS 5850, fan failed and they refused to send a replacement fan so it was sent to them for RMA at the same time as the ASUS, still haven't gotten that one back. I've had more XFX cards than all the other brands combined. I bought so many XFX strictly due to the double lifetime warranty (check the model, not all have the super warranty). Of those have had just 2 fan failures, both on 5850 cards. Both times contacted XFX support and they had complete new heatsink/fan assemblies at my door in 2-3 days, no cost or hassle to me. Also, my neighbor's running 4 XFX cards, no problems there either.

Regards/Beyond

Thanks for your opinion and the info about the types of problems. I have my GTX 570 running at 40C to 49C because it's in a cabinet I built that is air conditioned with cold winter air. In addition to that the GPU vents into a duct that carries the exhaust directly to either the outdoors in summer or into the furnace's cold air return duct in winter to heat the rest of the house rather than the cabinet and computer room. If I get a 7990 it will go into the same cabinet. I know I can sustain that operating temperature all summer, no problem, by patching in a small air conditioning unit that will supply the cold air mother nature so kindly provides now. It's a small AC unit and it needs to cool a very small volume so it will run very cheap. In my mind that solves the VRMs overheating problem.

The driver issues is something I cannot control so I'm going to avoid putting myself at the mercy of AMD or nVidia. I could wait until the card becomes obsolete before driver issues get fixed and I don't need that crap.

For now I'll buy just 1 X 7970 and put it in my host that doesn't have a GPU. I won't even need a new PSU to do that. Soon enough I'll be able to get a mobo with 2 PCI-E 3.0 x16 slots and put 2 GPUs in that.

Thanks for your opinion and the info about the types of problems. I have my GTX 570 running at 40C to 49C because it's in a cabinet I built that is air conditioned with cold winter air. In addition to that the GPU vents into a duct that carries the exhaust directly to either the outdoors in summer or into the furnace's cold air return duct in winter to heat the rest of the house rather than the cabinet and computer room. If I get a 7990 it will go into the same cabinet. I know I can sustain that operating temperature all summer, no problem, by patching in a small air conditioning unit that will supply the cold air mother nature so kindly provides now. It's a small AC unit and it needs to cool a very small volume so it will run very cheap. In my mind that solves the VRMs overheating problem.

Nice system. Mine is not so high tech, but works. In the winter the computers heat the whole (largish) house. I have them spread around strategically and the forced air fan distributes heat well. My furnace is combination wood and off peak electric and last winter was not used at all except for air distribution (and this is Minnesota). The electric company called and apparently didn't believe my explanation so sent out a repairman because they thought my off-peak electric meter must be broken. It wasn't. This winter (because of a couple cold snaps) have burned wood for parts of 3 days, no electric. In the summer I have a whole house fan that circulates a lot of air. Only used air conditioning two days last summer and that was because I felt sorry for Cocoa (see avatar).

Regards/Beyond

While I'm in a talkative mode: most of my PCs are running 2 GPUs per machine. For a year or two ran 2 boxes with 3 GPUs each. You know what, it was a PITA. Limitations in BOINC, more driver issues, higher temperatures. Not worth it for me. The 2 GPU boxes easily handle ANY summer heat and are quieter due to lower fan speeds. The best way IMO to run the 2 GPU boxes is with one ATI/AMD and 1 NVIDIA. That way no need for BOINC exclusions which are problematical due to BOINC bugs. With 1 of each it's easy to set a different project per GPU, no fuss, no muss. POEM is inefficient when trying to run it on more than one GPU per machine, but for instance running the ATI on POEM and the NVIDIA on GPUGrid works perfectly. Over and out...

Nice discussion and a pretty nice solution @ Beyond!

Regarding buying many more AMDs: I'd caution against that. While POEM runs great on them, POEM used to be short on work and may not recover quickly. If you run out of POEM WUs there are not too many attractive alternatives left. Mine runs Milkyway as a backup, not sure about Seti and Einstein. But all the "classical" projects AMDs could run since some time (Collatz, PG, Moo) I consider.. well, not very useful.

A single HD7970 won't hurt, though. Have fun with it.. it can either be a beast (~1.17 V, 1.2+ GHz) or run rather efficiently at 0.9 - 1.0 GHz with significantly less voltage :)

MrS

---

Scanning for our furry friends since Jan 2002

My "cooler cabinet" may sound high tech but it's not. I built 90% of it from scrap that cost me nothing; the remainder was about $20 worth of new clothes dryer vent duct. Even the paint I'm going to paint it with was scrounged for free. As soon as I paint it up pretty I'm going to disassemble it and make a video or series of photos while reassembling to show how I did it. The basic principle is... don't let the heat mix into the room and heat it up and then expect to use that hot air to cool stuff off, contain the heat the instant it comes out of the machine and deal with it sensibly. The alternative is monster, high dollar heat sink/fan combos or liquid cooling. All my gear works with stock fans and heat sinks. Like I keep telling mitrichr, brains (thinking), not money, is the answer. I'm not a brainiac, I just think a lot :-) Cuz I have no money!!

Your method of spreading the heat around the house is a good idea too. Very cute dog, btw :-)

Thanks for the tip on 1 AMD plus 1 nVidia per box. From various discussions I've lurked I kind of gathered that was a good way but never asked because never needed to. I'm going to plan my future purchases and expansions on that principle.

Over and out for me too, sorry for the thread hijack guys.

Nice discussion and a pretty nice solution @ Beyond!

Regarding buying many more AMDs: I'd caution against that. While POEM runs great on them, POEM used to be short on work and may not recover quickly. If you run out of POEM WUs there are not too many attractive alternatives left. Mine runs Milkyway as a backup, not sure about Seti and Einstein. But all the "classical" projects AMDs could run since some time (Collatz, PG, Moo) I consider.. well, not very useful.

A single HD7970 won't hurt, though. Have fun with it.. it can either be a beast (~1.17 V, 1.2+ GHz) or run rather efficiently at 0.9 - 1.0 GHz with significantly less voltage :)

MrS

Thanks! As far as projects (personally I like MW, Collatz, PG and Moo is OK too)(and don't forget Donate, the GPUGrid sister project), the tide is moving toward Open_CL. CAL is dead and both AMD and NVIDIA seem to be allocating more of their resources to the common language. Open_CL has recently made large performance strides on both platforms and is getting better with every release. It's just a matter of time IMO.

Regards/Beyond

Hi All,

We have no plans for an AMD app right now. As you may recall, we did have an OpenCL build ready some years ago, but the AMD drivers were never stable enough - and performance too low - for us to want to deploy it. Since then, we've worked further on the Nvidia application to improve its features and performance, and it would now be substantial effort to rejuvenate the AMD OpenCL code. Every time there's movement from the AMD camp, I do check it out to see whether things have improved, but realistically, it's unlikely that we'll make any change. Sorry to disappoint.

Plug: If you have an AMD GPU and really want to contribute to GPUGRID activities, we have our sister project http://donateathome.org/.[/url]

Dagorath:

If I could put a GPU on a Rasp-Pi I would do it, assuming there was the required app and other pre-requisites. Just enough CPU to drive a whopping GPU... that's the goal.

We build our own systems for our lab that are heavy on GPU and light on CPU. We also designed a custom case to deal with the cooling problems. I'll post more details if anyone's interested.

The cost of a single socket host system isn't too painful when amortised over 4 GPUs, and it's a lot less painful than making custom motherboards..

MJH