AMD R290/R290X
17 Dec 2013 13:21:23 UTC
Topic 197313
(moderation:
Did anybody try new AMD R290/X video cards for this project? Specification sounds as very tantalizing. My GTX660 Ti passed away to the computer haven so I'm looking for some excuse to order something like that.
AMD R290/R290X
)
Just read an article here regarding heating. You might give it a read.
The stock cooler is "barely
)
The stock cooler is "barely adequate" for these cards, as AMD would probably say. For noise sensitive people you can say: "the cards are completely unusable with the stock cooler". The issue becomes especially pronounced with Hawaii since the allowed clock speeds are hard-limited by the cooling solution, i.e. noise and gaming performance suffer from the inadequate cooling, which made AMD increase fan speeds after launch, mitigating the performance issue a bit. Even under GP-GPU benchmarks this thermal throtteling showed: the cards are often hardly faster than Tahiti on the far cheaper 280X!
However, none of this matters if you have a proper cooler. The best current air cooling solution seems to be the Accelero Extreme 3 with 4 additional memory heat sinks. There are others which also work very well, but none can match the price/performance of the Arctic Cooling.
Having said this.. I still don't know how the cards perform here and if this is worth the price premium.
MrS
Scanning for our furry friends since Jan 2002
I will wait with my decision
)
I will wait with my decision until 1.Q 2014
Maxwell, the next generation of nVidias GPU is said to be 4x faster / watt
https://dl.dropboxusercontent.com/u/50246791/nvidia-kepler-2012-maxwell-2014%2C4-1-299521-22.png
This specifically targets DP
)
This specifically targets DP performance, so the gain in SP is surely smaller. Could still be very impressive and significant, though. Yet I don't hold my breath for affordable Maxwells before middle of next year - 20 nm HP as apparently needed for large and fast GPUs hasn't even started yet. Or they could be using the low power process and "buy" this efficiency with significantly lower clock speeds - would be OK for some markets!
Waiting a bit more is also a good idea if you want a 290/X but don't want to modify the cooler yourself. The partner cards with custom cooling should be nearly here.
MrS
Scanning for our furry friends since Jan 2002
I asked on the Milkyway
)
I asked on the Milkyway forums about the R9-290x. 288larsson was kind enough to test out the R9-290x card with Einstein.
http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=3382
Unfortunately, the R9-290x is running a bit slower than a 7970 with this project. I think the 7970 or the new R9-280x would give the best price to performance ratio for this project.
RE: I asked on the Milkyway
)
Thank you for the link. The information is very valuable for me.
Short excerpt from discussion:
Hello,
I was wondering if you could test out your R9-290x with running three Einstein BRP5 tasks (0.33 utilization factor) and post what the run time is per task? If you could run a batch of GPU tasks without any other CPU tasks running that would be great.
Thanks,
Jeroen
Hi 8390sec Catalyst 13.11betav8
Here is a second data point
)
Here is a second data point for the new AMD R9-290X GPU compared to my previous stock clocked 7970. The R9-290X is 6 to 8 percent slower than the 7970 for Einstein GPU jobs.
Details, I purchased the Gigabyte custom cooled R9-290X to avoid all the thermal throttling problems the reference cards were having. The good news is that the custom cooling solution works and the card maintained its 1040 MHz core clock speed during the past two weeks of 24/7 testing. I used the MSI Afterburner program to track the frequencies and temperatures. Driver version 13.12 was used.
I tested 1, 2 and 4 jobs GPU jobs (Arecibo and Perseus mixed and homogenous) running simultaneously while always maintaining enough CPU free to insure the GPU was fed properly. I turned hyper-threading off and ran CPU jobs on the free cores. I also testing running only GPU jobs and I saw the same results against the 7970.
The supporting platform is top notch, i7-4770K clocked at 4.2 GHz, 2 x 4 GB 2400 DDR3 RAM, PCIe x16 version 3 with Windows 7 64 bit. So I don’t believe the CPU support in any way inhibited the GPU.
So at present there is no reason to use a R9-290X over the 7970 for Einstein GPU jobs. I’ll keep an eye on driver updates and report back if there is any big changes.
Thanks for the detailed
)
Thanks for the detailed information Robert!
It's surprising, since one might expect the Hawaii-XT to perform better in single-precision (when properly cooled) than the Tahiti-XT because of the higher shader count.
Thanks for the info. Just
)
Thanks for the info. Just speculating, could be that the Einstein application does not use all the shaders? or could the developers add some tweaking to increase performance in 290(x) cards?
I was also expecting better performance provided the higher count of shaders vs 7970.
Hawaii performaning worse
)
Hawaii performaning worse than Tahiti at Milkyway is expected since it drops DP performance down to 1/8 SP, coming from 1/4 SP in Tahiti. However, the lack of performance improvement at Einstein surprises me. Ruling out the CPU (Einstein CPU load is also low for AMD GPU tasks, right?) this leaves the PCIe bus possibly being saturated or the memory bandwidth limiting, unless there are stranger issues.
What about GPU load, running the same number of tasks in parallel? Is Hawaii less loaded than Tahiti? What about task scaling? Maybe Hawaii needs more in parallel? Or 3 might be the sweet spot? (probably not since scaling is often pretty flat at 3 or 4 tasks, without much of a penalty for more)
Does GPU memory over/underclock strongly influence results (performance and GPU load)?
Anyway, thanks for your feedback so far, Robert! My questions here are just a few points I'd investigate if it was my card. No worries if you've got better things to do or already checked some of them :)
MrS
Scanning for our furry friends since Jan 2002