Message boards : Graphics cards (GPUs) : Most cost effective setup?
Author | Message |
---|---|
I currently have a RAC of 292k+ with a single GTX 660, running on Linux. The GPU was under $200 CAD, and I am currently paying 6.9 cents (CAD) per kWh. I haven't measured the draw at the wall, but I feel like this is a pretty good setup. Just curious :) | |
ID: 33564 | Rating: 0 | rate: / Reply Quote | |
Yes it is, i'm still trying to get things figured out in terms of performance per dollar/watt upgrading my old C2D/gt 240/430 rigs. My main rig had a GTX460 and i'm also looking at either a 660 or 660Ti since i heard they were getting even more price drops in November. | |
ID: 33566 | Rating: 0 | rate: / Reply Quote | |
I'm interested in this too. My current rig is a GTX460 1GB @ 800MHz. Thinking of upgrading to either 660 / 660Ti / 750Ti (assuming it is a 660 with 256bit memory and higher clocks). | |
ID: 33567 | Rating: 0 | rate: / Reply Quote | |
Due to the 'MSU overhead' (Motherboard, CPU, RAM, Drives, fans), in terms of credits/Watt the best systems have multiple GPU's. | |
ID: 33570 | Rating: 0 | rate: / Reply Quote | |
Hi, | |
ID: 33575 | Rating: 0 | rate: / Reply Quote | |
If you really wanted to keep to 100W you would need to get a GTX650 (65W TDP), and run it in a mini-itx setup. Perhaps an Atom (though Bay Trail might not show to Feb/Apr '14), AMD E2-1800 (18W TDP), J1850 Celeron (10W), J2850 Pentium (10W) but possibly even up to an Intel i3-3220T 3rd Gen Core 2.6GHz LV Processor (35W TDP) so long as you didn't crunch on it for CPU work. | |
ID: 33579 | Rating: 0 | rate: / Reply Quote | |
Skgiven if you run it for 2 years you would get more credits I think...calculation error? | |
ID: 33580 | Rating: 0 | rate: / Reply Quote | |
Thanks for the responses. I'm using my CPU for WGC projects, while the (single) GPU runs GPUGrid. The setup is using hyperthreading, so it doesn't consume a whole CPU core. I suppose my view of efficiency is to find an efficient GPU and put it to work, rather than a dedicated multi-GPU setup. | |
ID: 33581 | Rating: 0 | rate: / Reply Quote | |
I should clarify that 7 cents is the base rate in British Columbia. With fees and taxes it comes to about 9 cents per kWh. | |
ID: 33585 | Rating: 0 | rate: / Reply Quote | |
My G2020 supports a GTX670 and a GTX650TiBoost and is presently drawing 315W (240V), also Linux. Two GTX660's would have the same crunching performance, which roughly matches a Titan. The power usage of two GTX660's is going to be around about the same as a 670 and 650TiBoost, but your AMD FX(tm)-6300 has a TDP of 95W, whereas my G2020's TDP is only 55W, and is probably not drawing more than 30W. My guess is your system with an extra GTX660 would use around 350 to 380W, assuming you are not overclocking, not crunching CPU projects (which you dont appear to do) and you keep the power usage in check by having good cooling. Your theoretical RAC would also be around 660K/day, but you're probably going to get the odd failure that would bring it down to around 600K/day. | |
ID: 33590 | Rating: 0 | rate: / Reply Quote | |
Throw in my tuppance as running a budget machine: | |
ID: 33604 | Rating: 0 | rate: / Reply Quote | |
I agree with skgiven, that a single GPU doesn't take advantage of the system overhead enough. I plan to add a second GTX 660 at some point and hopefully will get close to a 600k RAC. This will draw maybe between 400-450W at load? I run two GTX 660s on an Ivy Bridge i7-3770 board, with two virtual cores devoted to the GPUs and the other six running CEP2 on World Community Grid. It draws 345 watts from the plug, using a Seasonic bronze 85+ power supply. (These days, I buy the Rosewill Gold 90+ supplies from Newegg, if you can get them in Canada.) It is not overclocked, and I think I was getting around 550k RAC, maybe a little more, but just started up again and don't have the statistics yet. | |
ID: 33605 | Rating: 0 | rate: / Reply Quote | |
I'm curious about CPU power consumption when not using all of the cores. My AMD FX-6300 has 3 cores (6 virtual), and my Kepler GPU will keep one thread going. I assume the consumption is higher than simply dividing 95W by 3. As a guess, if it is using 40W with one GPU and maybe 70W with two GPUs, then the gap may not be as big versus a G2020 utilizing its 2 cores? | |
ID: 33607 | Rating: 0 | rate: / Reply Quote | |
two virtual cores devoted to the GPUs and the other six running CEP2 on World Community Grid. Interesting... was the optimum to use two HT cores for the GPUs? Would you get more total output if you dedicated just 1 HT core to the GPUs and 7 to CEP2 (WGC)? | |
ID: 33608 | Rating: 0 | rate: / Reply Quote | |
Interesting... was the optimum to use two HT cores for the GPUs? Would you get more total output if you dedicated just 1 HT core to the GPUs and 7 to CEP2 (WGC)? I take a hit in output if I use only a single CPU core for both, though at the moment I don't remember how bad it is. I think it depends on the type of work unit; some can live with only half a virtual core, and those don't fare so badly, but even there you get a small loss. Since the GPU is the higher-output component, I like to keep it fed properly. By the way, I think the Seasonic is actually called an 80+, though it runs at about 85 percent efficiency at my load (it is rated at 550 watts). | |
ID: 33610 | Rating: 0 | rate: / Reply Quote | |
There are 80+ bronze, gold and platinum ones. And reading through all the specs of several brands, it seems that the platinum is to most efficient and the bronze the less. | |
ID: 33614 | Rating: 0 | rate: / Reply Quote | |
There are 80+ bronze, gold and platinum ones. And reading through all the specs of several brands, it seems that the platinum is to most efficient and the bronze the less. True. The Golds are now usually around 90% efficient over most of their range, and the platinums are slightly higher insofar as their rated minimum values are concerned. But in practice, the last time I was shopping, the Rosewill Golds were in reality as good as the Platinums over most of their range, and there was no point in paying the cost difference. It depends on the brand of course. But remember that increasing the efficiency by only 1 percent (i.e., from 90% to 91%) decreases the heat that the power supply has to dissipate by about 10%, so its fan will run less and can be quieter. Also, the components will tend to last longer. And a good brand will actually deliver at it rated value, unlike lesser brands. I often see people buying ridiculously large power supplies to supply fairly ordinary upper mid-range cards, which should not be necessary at all. | |
ID: 33615 | Rating: 0 | rate: / Reply Quote | |
I'm running 2 x GTX 570 and a OC'd i5 4670k with a 550W be quiet! psu at the moment. | |
ID: 33618 | Rating: 0 | rate: / Reply Quote | |
Interesting... was the optimum to use two HT cores for the GPUs? Would you get more total output if you dedicated just 1 HT core to the GPUs and 7 to CEP2 (WGC)?I take a hit in output if I use only a single CPU core for both, though at the moment I don't remember how bad it is. I think it depends on the type of work unit; some can live with only half a virtual core, and those don't fare so badly, but even there you get a small loss. Since the GPU is the higher-output component, I like to keep it fed properly. My i7-3770K (still @4.2GHz) supports 3 Kepler GPU's (for which the GPUGrid apps dedicate a full CPU/thread), a GTX 770, 660Ti and 660. With a 70% CPU usage configuration, one of several reasonably happy mediums (typically 49% to 76%), Boinc runs 4 CPU tasks and 3 GPU tasks. The CPU usage is around 90% meaning that 7 of the 8 CPU threads are fully used and the 8th thread is largely available (but gets some use). However, even with this setup my 770 struggles a bit; GPU usage is only ~74%. The 660 is 91% and the 660Ti is ~87%. The 770 struggles the most as it's the best GPU of the three and the most reliant on the CPU, but this depends on the CPU projects and the GPU task types being run (utilization is not always the same). When I set CPU usage in Boinc to 51% Boinc runs 3 CPU tasks and 3 GPU tasks, uses just over 75% of the CPU (6/8), GPU usages are: GTX770 ~80% GTX660Ti ~93% GTX660 ~89% On the face of it that's only a 6%, 2% and 2% difference. However, runtimes tend to show a greater improvement. When I let Boinc run at 100% CPU usage (with default GPUGrid/Boinc setup) my WU's say they will use ~0.59 CPU cores per GPUGrid task (as read by scheduler). The scheduler thinks this is ~1.77 CPU's, and being less than two logical CPU cores (threads in my case) will allow 7 CPU tasks to run as well as the 3 GPU tasks. With such a BAD setup my GPU utilization fell to on average ~55% but also very spiky (dropping to 30% at times and occasionally peaking >80%). Also worth noting is that some CPU projects suffer extremely badly with this setup, it can impact badly on VM's and force Boinc into panic mode... Clocks also drop - my GTX660Ti drops from 1189/1202 to 1058MHz and my 770 sometimes drops from 1202 to 1045. My 660 for some reason remained oblivious to just about everything and stayed exactly at 1071MHz. When I only run GPUGrid work my GPU's usage rises the most. My GTX770's usage goes up to 92% and my GTX660 rises to 96%. For the GTX770 optimal is 92% faster than with the CPU usage set to 100%. Unfortunately, we just can't have it all our own way - utilize all resources to 100%, there is always contention. If you chose to utilize towards your CPU that's up to you, but I tend to utilize towards my GPU's as they do more work and cost a lot to buy and run. The best setup for GPUGrid is 4 Titans supported by an overclocked i7 with HT off on Linux and only running GPUGrid work. Anything close to that is a great setup. ____________ FAQ's HOW TO: - Opt out of Beta Tests - Ask for Help | |
ID: 33620 | Rating: 0 | rate: / Reply Quote | |
I'm running 2 x GTX 570 and a OC'd i5 4670k with a 550W be quiet! psu at the moment. If your system is pulling 500W on a 550W PSU, it will run less efficient, it will run hotter, with a shorter lifespan and it may result in some errors of GPU WU's. The advice is to use a PSU with plenty of head room. If it works at 50-70%, so a 1000-850W PSU in your case, is the best for performance, efficiency and heat production. ____________ Greetings from TJ | |
ID: 33626 | Rating: 0 | rate: / Reply Quote | |
450W from a 550W PSU might be pushing it a bit, but it really depends on the PSU, some are perfectly capable of supplying >80% of their maximum load continuously and efficiently, others are not. | |
ID: 33633 | Rating: 0 | rate: / Reply Quote | |
Nice PSU graphs. I have an HX850, so I feel it should be pretty efficient when I get two GTX 660s in there. It looks like a target of 50% is great for budget PSUs, but a high end PSU like the AX860i would more easily handle additional load. | |
ID: 33643 | Rating: 0 | rate: / Reply Quote | |
Hm ok does anything changed in technology? I heared years ago, the psus with 80.83 85 certificate are running only at this rate in the upper usage. So i have as example buyed an 530w psu for my dual 570gtx setup here. The other psus in other machines dont have much headroom either. But as i read here this is not the case anymore, good to know O.o | |
ID: 33644 | Rating: 0 | rate: / Reply Quote | |
I'm running 2 x GTX 570 and a OC'd i5 4670k with a 550W be quiet! psu at the moment. It's not a 24/7 crunching machine, so it doesn't matter. 500W is the peak with 100% and both GPU's at 99%. I'm only doing GPU WU's every few days and while crunching with the CPU it draws ~ 150W. | |
ID: 33645 | Rating: 0 | rate: / Reply Quote | |
Looking at those PSU efficiency charts, we're looking at around 3% difference in efficiency comparing between the peak ~45% load and at 90% load. | |
ID: 33650 | Rating: 0 | rate: / Reply Quote | |
Example of my four years old PSU, 20-50% mean the range between 130-310 watts. Buying a new one with 90% (4-6% better than now) starts at 75 Euro (PSU with same net power) or 300 kWh (0,25 Euro per kWh). With 5% increased efficiency that is according to my current wattage level (~170 watts) 170*5% = 161 or ~10 watts. Price of 300kWh to reduce 10 watts requires a minimum runtime of 30.000 hours or (current daily runtime in my case max. 14 hours) 2143 days or 5.8 years. It´s too costly than it´s worth. Maybe I should switch to a platinum PSU. Or simply reduce the DC runtime at all. | |
ID: 33654 | Rating: 0 | rate: / Reply Quote | |
For a good crunching system (say with 2 high end GPU's or 3 mid-high end GPU's) with a power draw of say 500W, you would notice the difference more. The more energy efficient PSU would result in a cooler system, and might save you money in the long run. | |
ID: 33664 | Rating: 0 | rate: / Reply Quote | |
Message boards : Graphics cards (GPUs) : Most cost effective setup?