Comprehensive GPU performance list?
10 Jan 2018 18:52:08 UTC
Topic 212584
(moderation:
I have looked around the site a bit and I have not been able to find a comprehensive list of GPUs that compares their performance in E@H.
If it exists, I am looking for a list that compares both nVidia and AMD GPUs from recent generations and gives the relative performance of both in E@H
I am specifically interested in the AMD 500 (580) and Vega series cards and the nVidia 900/1000 series cards particularly the X60s or above. I already know that the AMD cards consume more power than the nVidia cards, but that is not a concern for me.
Does such a list exist?
If not, does anyone know if either AMD or nVidia are the better performing cards?
If anyone has any input, I would appreciate it.
Thanks.
Wiyosaya wrote:I have looked
)
Such a list doesn't exist. By searching in the forums, you may find some anecdotal evidence that could allow you to make fairly general conclusions. A small number of users have published stats about their own machines but these would be rather dated now. Your only real option, other than hoping that people will comment on their own current findings/preferences (which you should try to confirm independently), is to spend time going through the top hosts list looking for particular models that interest you.
It's easier if there is only a single GPU. It gets pretty messy for multi-GPU scenarios. It also gets tricky in correctly guessing the task concurrency of any particular GPU. One could be running 3 at a time and another only two. You also have to be careful with multi-GPU hosts because BOINC may not correctly list each type. It could even look like they're all the one type when they're actually not.
Does initial capital outlay bother you? Both concern me (since I pay for both) but I've gone with AMD in recent times because nvidia cards are a lot more expensive in this country. The most recent nvidia card I have is a 750Ti and it performs at about one third of an old HD7870 I'm also running. I don't usually play with high end cards but I believe (from what others have written) that high end stuff like 1070, 1080, 1080Ti are very productive and quite cheap on power. They just cost a ridiculous amount here. I've bought a couple of RX 570s recently quite cheaply and am very happy with the performance - RAC ultimately around 560K. Each card has been an upgrade to an old CPU only cruncher - eg Q6600 quad core from 2008 :-).
Cheers,
Gary.
This would be a very
)
This would be a very difficult task because a lot of people run multiple projects. I recently looked around at people's stats and the hardware/OS configurations they were using and the results were all over the place. Some people pause work to play games, or have other downtime that confounds analysis without participation.
Unfortunately the operating system you use plays into this as well. There seems to be differences between Windows and Linux for various cards, and while macOS keeps pace on the AMD side, the Nvidia work units perform abysmally on macOS here.
I originally blamed Apple's outdated OpenCL, or Nvidia's drivers, but found they work well in other benchmarks and other BOINC projects. So it must be a difference in the software here (note that all other OS's have FGRPopencl1K-nvidia). I have a machine that dual boots Win10 and macOS and the Nvidia GPU performs almost twice as well under Win 10.
But if any of you want to put something like this together, count me in. I've been researching GPU upgrades too and have collected some data that may help.
GPUGRID has compiled a graph
)
GPUGRID has compiled a graph of GPU performance on their web site (at the bottom of the page) - but that is the project. Even in this project's case, with being able to run more than one WU on a GPU, I tend to agree it would likely be a difficult task unless the information on how many WUs are running at a time on a GPU is available. I have started running more than one WU on my systems for this project, and the metric I am using is WUs processed in 24hrs.
It is the initial capital outlay that is my primary concern, however, I do pay for both. I just cannot see the difference in energy consumption even running 24/7 making up for the difference in price.
If you are willing to reveal it, I am interested in the specific stats on those 570s. Is the 560K RAC per card and how many WUs are you running per card, are my specific questions. I currently have a 6GB 980 Ti I picked up used, and a 3GB 1060 that I am running. The 980 Ti is in one PC, currently running 6 WUs and trying to find a sweet spot for CPU usage - its RAC is 490K. The 1060 is running on another PC 3 WUs at time and trying to find a sweet spot for CPU, too. Its RAC is 290K.
https://setiathome.berkeley.e
)
https://setiathome.berkeley.edu/gpu_list.php
The 1060 is running on
)
Wiyosaya wrote:... I tend to
)
Concurrency (C) is fairly easy to estimate if a machine runs a single project 24/7 and if its RAC is close to the plateau and the owner isn't disturbing things :-). If you take an average for the elapsed time (E) using say 20 recent tasks from the website, then the theoretical tasks/day (T) can be expressed as
T = 86400 x C / E or, rearranging, C = T x E / 86400To eliminate T, the other relationship needed is the equation for theoretical RAC (neglecting any CPU tasks)
RAC = T x 3465 or, rearranging, T = RAC / 3465Eliminating T, by combining the two equations, you get
C = RAC x E / 3465 / 86400 or (what I tend to use occasionally) C = RAC(K) x E / 299376This host is an old Pentium dual core that now has an RX 570 upgrade. Its theoretical RAC is 560K (actual is 545K and it's still rising slowly) and when I first monitored it after it got going, it was averaging around 1620 secs per task. So you can tell me how many concurrent tasks it runs :-).
On the top hosts lists on the website, the current #15 (owned by Gavin) has a single Vega and has a current RAC around 1200K. The most recent page of validated tasks shows elapsed times between 460 and 490 seconds so I called it 475 without adding them all up. So, concurrency = 1200x475/299376=1.904, so fairly obviously it would likely be 2. If so, and if 475 is a reasonable average, you can work back to find what is likely to happen to RAC if everything remains stable. Theoretical RAC = 2x299376/475=1260K.
As a general guide, if the calculation for concurrency gives an answer that's close, but on the low side, the current RAC should still be on a rising trend. The key thing is to pick hosts that are known to be stable - ie you need to observe them for a while :-). If the calculation is a long way from an integer answer, check back through pages of results to see any causes. There might be a recent concurrency change or signs of non-24/7 running. You just have to move on to a different example if things don't seem to be stable. You can rely a bit on human nature. If a person is keen enough to have a host on the top hosts page, chances are they really, really want to keep it there and get it as close as possible to the top :-). So, it's very likely to be running 24/7 and as efficiently as possible and not being fiddled with unnecessarily :-). For useful stats purposes, it would be very nice to have an alternative top hosts page, limited to machines with a single GPU :-). That would make it a lot easier to come up with a meaningful GPU comparison list based on EAH performance. Gavin could be #1 and #2 on that list :-).
It depends on how long term you are thinking. If you are in it for the long run, it's surprising how a little bit of energy efficiency can add up to a decent amount of savings over time. AMD has improved (I don't know about Vega) in that I'm running RX 460s and 560s that don't need PCIe power cables and are quite miserly in what they use for crunching. As an example, I have an RX 460 in an old E6300 Pentium dual core host (2009 vintage) using a 300W PSU that was 80% efficient (although not certified) when new in 2006. The OEM was SeaSonic and it's rated at 270W on the 12V rail and I did a cap replacement about 2 years ago. It's pulling about 140W from the wall (the display fluctuates between about 125W and 149W quite regularly). The RAC is about 270K and the average task completion time is around 2200 secs. So you can tell me how many concurrent tasks I'm running :-). As well as the GPU tasks, it does crunch a single CPU task.
I haven't done a proper power analysis on a 570 yet. I did have a quick look at one point and I seem to remember around 270-280W or so when crunching. I need to go and do it properly. My gut feeling is that it's pretty comparable - in other words it uses twice as much power and produces twice as much output when compared to a 460. Seeing as I had a lot of older CPU only hosts and a lot of 300W PSUs in them, and the 460s cost quite a bit less than half of what I paid for the 570 (I got them as throwout specials), my feeling is the 460 was a better investment for me.
Cheers,
Gary.
http://www.geeks3d.com/201403
)
http://www.geeks3d.com/20140305/amd-radeon-and-nvidia-geforce-fp32-fp64-gflops-table-computing/
Gary Roberts wrote:For useful
)
I second this motion!
Having an alternative list of top hosts limited to one GPU would be very informative indeed. It's maybe a big ask but it would be great if the dev's could at least find a little time to investigate making this happen. So many repeated questions from so many users could be addressed with such a resource and I could be (temporarily) undisputed King of the castle :-)
I know that most participants do not have or run multi GPU behemoth computers and from my own experience it is frustrating not knowing how your particular single card system and strong commitment to the project compares to others on a 'like for like' basis due to the fact that the current Top Host list is dominated by machines that are in many cases far from average... Whales in a sea of minnows.
I for one do not see a downside to this suggestion, encouraging competition and displaying achievable goals will by default increase project productivity and audience participation and that means everybody wins!
Gav.
Is there a way to scrape the
)
Is there a way to scrape the database that feeds the stat page so we can do that ourselves? I'm guessing not because of the option to hide or show your computers, and some may not take kindly to their hardware and software configuration getting exposed if they elected to hide them.
I'd at least love the option to do this myself (aside from manually copying my stats pages to analyze the performance data).
Didn't some projects list
)
Didn't some projects list credit by OS and CPU type once upon a time? It became a joke though as every variant of Windows and every other OS got broken into separate categories, and then really got out of hand with people apparently spoofing the OS and CPU identification.
But it would be nice to search for all of the stats for a particular GPU type under various OSs, and this could be done without tying the configuration back to a particular user to respect their privacy.