WCG GPU Crunching: Performance/Value

[xvi]

What/Why?
I've spent some time trying to research GPU crunching performance and had made myself a chart to help visualize this. I pulled the only somewhat trustworthy performance numbers I could find and tossed a few simple functions on it to breakdown things like raw performance and value. The results are the following.


Raw Performance
First is a raw "more is better" chart based on the inverse of of the average time to complete a WU in seconds. Given that some of the names provided are nondescript, I've made assumptions based on the given performance. For example, two "Radeon HD 7900 Series" cards with obvious steps in performance are assumed to be the HD 7970 and the HD 7950. This will be verified/updated when user-submitted results come in.

Price per Performance (Value)
If something's not good value, it just rubs me the wrong way. This chart takes GPU performance and divides that by the street value as determined by retail price for latest-generation GPUs or eBay price for older GPUs. This chart answers the question "How many points can I get for X amount of dollars?"
As stated above, given that some of the names provided are nondescript, I've made assumptions based on the given performance.
(The chart will be put back in order in the next update.)

Help improve the charts!
If you'd like to contribute, I'd appreciate submissions. You can find GPU your GPU Work Unit completion time here as long as you know the result name that correlates with your GPU. You're looking for elapsed time in hours.
Please fill out the following:

WCG Username: <If different than forum>
CPU Model/Speed:
GPU 0 Model:
GPU 0 Time per WU:
GPU 0 Clock Speed: <If overclocked>
GPU 0 Work Unit Completion Time:
GPU 0 GPU Result Name: <optional>
Add additional GPUs as needed

To Do List:

Spoiler

• Update prices for all cards
• Come up with better pricing strategy for old cards (Worthmonkey not accurate, eBay fluctuates)
• Add performance per watt charts (just declared TDP?)
• Research/Create CPU charts
• Fix price/performance graph
• Make charts prettier

Changelog:

Spoiler

2/13/13: Initial release! WYSIWYG!

[xvi]

<Reserved>

[xvi]

<Also Reserved>

[[Ion]]

I appreciate the work that you've put into this, but I must say I'm a bit surprised/suspicious of some of the results. You have the HD7950 doing 1.85x the output of the GTX470, but I can state from personal experience that's not true. My GTX470s are each slower than a HD7770, and my system with two GTX470s does just under half what my single-HD7950 setup did.

This is a really cool idea!

[Norton]

Pretty cool setup!

FYI- the Radeon (unknown) is likely the 7850/70's iirc my 7870's show up as unknown and they should fall into that place on the chart- they both do around 70k ppd each

[xvi]

This started as a "I want a new video card. What should I buy?" sort of thing and just took off a little more than I expected.

Originally Posted by [Ion]

I appreciate the work that you've put into this, but I must say I'm a bit surprised/suspicious of some of the results. You have the HD7950 doing 1.85x the output of the GTX470, but I can state from personal experience that's not true. My GTX470s are each slower than a HD7770, and my system with two GTX470s does just under half what my single-HD7950 setup did.

This is a really cool idea!

All performance numbers were pulled from a post from a self-proclaimed WCG tech on the WCG forums completing unknown work units on an unknown client with an unknown config. I'm using those numbers because I can't find anything else as complete. The only GPU I know is correct is a yet-to-be-added-to-the-chart HD 5770 which comes out to roughly the same performance as the "Radeon 5700 Series (Juniper)". Based on that, I assumed all the other results were roughly accurate. I figure they should be a decent placeholder until I get some more verifiable results.

While I will admit to preferring AMD over nVidia, I wouldn't intentionally skew results either way. I'll happily take donations of cards for independent (and indefinite) testing though.

Originally Posted by Norton

Pretty cool setup!
FYI- the Radeon (unknown) is likely the 7850/70's iirc my 7870's show up as unknown and they should fall into that place on the chart- they both do around 70k ppd each

That's enough proof for me. I'll add it. Keep in mind that "XVI's Magical Performance Units" aren't currently based on points whatsoever (although they might be in the future).

[[Ion]]

Oh, I'm certainly not accusing you of biasing the results. I'm just curious where the numbers came from and how they were calculated, that's all

[xvi]

Originally Posted by [Ion]

Oh, I'm certainly not accusing you of biasing the results. I'm just curious where the numbers came from and how they were calculated, that's all

Like I say, just the first thing I managed to grab. I'm not sure how consistent the results will be across different work units and especially across different WCG projects, so I think I'll have to convert to a PPD-based scoring. I just haven't quite figured out an easy way to find the PPD of just the GPU.

[Norton]

Originally Posted by xvi

Like I say, just the first thing I managed to grab. I'm not sure how consistent the results will be across different work units and especially across different WCG projects, so I think I'll have to convert to a PPD-based scoring. I just haven't quite figured out an easy way to find the PPD of just the GPU.

I don't know about the other generations or NVidia but the Radeon 7xxx series seems to scale pretty close to linear based on # of stream processors... gpu speed, memory size, cpu type/speed have a lesser impact.

Helpful for your project? Dunno

[mstenholm]

You input template does not take into consideration that a GPU can do 1, 2, 3...16 WU at the time. Personally I prefer number of WU in a day taken over a few days to even out weekend fluctuations in amount of PVs. The original data was compiled before the WU got double in size and people in general found out that they could run more then one at the time.

My data:

7970 @ 1045 MHz 1900/day (12 WU)
7950 @ 1100 MHz 1600/day (10 WU)
7770 @ 1100 MHz 730/day (5 WU)

[james888]

This is interesting. You might want to do a performance/watt graph also. Edit: Ah that is in the to do list. Maybe you should use wizz's reviews to find watts used. Maybe just use reference.


WCG Username: <jjames888>
CPU Model/Speed: 2500k 4.4ghz
GPU 0 Model: 7970 (16 WU's)
GPU 0 Time per WU: ?
GPU 0 Clock Speed: <1200/1600>
GPU 0 Work Unit Completion Time: .2 hours (720 seconds, or 12 minutes) I do range from .16 - .22 hours elapsed time (about 10-13 minutes)
GPU 0 GPU Result Name: X0930110551170200907070826

[Chevalr1c]

WCG Username: Chevalr1c
CPU Model/Speed: Core 2 Duo E8400
GPU 0 Model: Radeon HD 6670 DDR3 (Turks XT)
GPU 0 Time per WU: 0.04 - 0.05 hours (I assumed that the "CPU time" values have to be used here)
GPU 0 Clock Speed: <If overclocked>
GPU 0 Work Unit Completion Time: 0.18 - 0.24 hours (I used "elapsed time" here)
GPU 0 GPU Result Name: I looked a a batch of 3 pages on the "Results status" page and noted minima and maxima.

[brandonwh64]

what is AMD Radeon HD (Unknown)? Since it was third on the list for higher PPD output is what made me curious.

[ChristTheGreat]

so HD6950 does a better job than the GTX 680 xD

[Norton]

Originally Posted by brandonwh64

what is AMD Radeon HD (Unknown)? Since it was third on the list for higher PPD output is what made me curious.

At the time the original list was made it was the Radeon 7850/7870 (the older BOINC Manager couldn't pick up the specs on the card). The 7.0.40 and up revisions seem to pick up the specs my 7870 just fine now.

[JNUKZ]

Looks like any HD7xxx are a beast even my 7770 is better than GTX 680.

[xvi]

Originally Posted by mstenholm

You input template does not take into consideration that a GPU can do 1, 2, 3...16 WU at the time. Personally I prefer number of WU in a day taken over a few days to even out weekend fluctuations in amount of PVs. The original data was compiled before the WU got double in size and people in general found out that they could run more then one at the time.

My data:

7970 @ 1045 MHz 1900/day (12 WU)
7950 @ 1100 MHz 1600/day (10 WU)
7770 @ 1100 MHz 730/day (5 WU)

The only card I've managed to test is a HD 5770 and it seemed to match the original data somewhat closely. The (admittedly old, but statistically sound) data I've used is more of a placeholder until I can get some more recent results. Again, I think a system based on PPD would be better, but it's harder to implement with WCG than it is with Folding@Home.

You're right about the multiple WUs thing though. Unfortunately for this comparison, yes, multiple running multiple WUs at once can get more performance out of a card, but it makes a royal mess of things when trying to compare cards. Again, working off of PPD would help even this out, but it would still create some dirty data until a good number of results came in and everything can average out. Unless I can reliably calculate the increase in performance divided by the number of WUs running at once, results running multiple WUs might have to wait. I suppose the big question is whether or not Card X benefits more from that than Card Y, in which case single WU results should be thrown out completely and multi-WU results should be used.

I guess I'm forced to buy a new video card so I can test this out. Oh the drudgery.

Originally Posted by james888

This is interesting. You might want to do a performance/watt graph also. Edit: Ah that is in the to do list. Maybe you should use wizz's reviews to find watts used. Maybe just use reference.

Yep! Already in the to-do list. I'd very much prefer actual power consumption rather than what it says on the box. Wizzard does a pretty good job of calculating consumption as does Anandtech. Sadly, the power draw will be under full gaming and not actual crunching, but it should be a very good indicator regardless. All of this is in the works though.

WCG Username: <jjames888>
CPU Model/Speed: 2500k 4.4ghz
GPU 0 Model: 7970 (16 WU's)
GPU 0 Time per WU: ?
GPU 0 Clock Speed: <1200/1600>
GPU 0 Work Unit Completion Time: 720 seconds (12 minutes)
GPU 0 GPU Result Name: X0930110551170200907070826

Originally Posted by Chevalr1c

WCG Username: Chevalr1c
CPU Model/Speed: Core 2 Duo E8400
GPU 0 Model: Radeon HD 6670 DDR3 (Turks XT)
GPU 0 Time per WU: 0.04 - 0.05 hours (I assumed that the "CPU time" values have to be used here)
GPU 0 Clock Speed: <If overclocked>
GPU 0 Work Unit Completion Time: 0.18 - 0.24 hours (I used "elapsed time" here)
GPU 0 GPU Result Name: I looked a a batch of 3 pages on the "Results status" page and noted minima and maxima.

Dangit, you're right. "GPU 0 Time per WU" and "GPU 0 Work Unit Completion Time" are supposed to be the same thing. I was obviously asleep while writing that up.

In my testing, "Elapsed time" was what correlated with the results I've scavenged, but "CPU time" would probably be more accurate while running multiple units. I assume by "CPU time" they mean "GPU time". If not, my scavenged results will not (easily) compare with new results.

Originally Posted by Norton

About 20 seconds longer than single wu's (appr. 01:45 vs 01:25 to 99.415%)

This still true, Norton? 1WU at 85 sec vs 3WU at 105 sec (/3 = 35 sec per WU?). That's a 243% increase. o.0

[mstenholm]

At one point I did some timing of different numbers of WUs on the same rig:

7770 - 3 each took 6:55, 4 took 8:31 each
7950 - 4 each took 4:34, 6 took 6:25 each and I ended up with 10 since it produced the highest points wise

I think that you should stay with your original data set and just normalize them with 7970 being 100 % and then add that you can run multiple WU. We can input the most used number : i.e 7970 is 12 but the range is 10-16. The optimal number will differ from rig to rig so a little trial and error is required. Best of luck with whatever solution you chose.

[AnnCore]

So you got me curious too, and I recorded a few results...

This is with 4 WUs: 2 CPU WUs and 2 GPU WUs (each GPU: 1CPU + 1 ATI GPU)

Note: No screensaver.

WCG Username: AnnCore
CPU Model/Speed: QX9770
GPU 0 Model: ATI Radeon HD 5850
GPU 0 Time per WU: 8:36
GPU 0 Clock Speed: <If overclocked>
GPU 0 GPU Result Name: <optional>

GPU 1 Model: ATI Radeon HD 5850
GPU 1 Time per WU: 8:45
GPU 1 Clock Speed: <If overclocked>
GPU 1 GPU Result Name: <optional>

And then this is with 4 WUs: 2 CPU WUs and 2 GPU WUs (each GPU: 1CPU + 1 ATI GPU)

Note: With screensaver.

WCG Username: AnnCore
CPU Model/Speed: QX9770
GPU 0 Model: ATI Radeon HD 5850
GPU 0 Time per WU: 9:17
GPU 0 Clock Speed: <If overclocked>
GPU 0 GPU Result Name: <optional>

GPU 1 Model: ATI Radeon HD 5850
GPU 1 Time per WU: 9:22
GPU 1 Clock Speed: <If overclocked>
GPU 1 GPU Result Name: <optional>

Last I tried only GPU WUs (1.0 CPU + 0.5 ATI GPU)

This is with 4 WUs: 2 CPU WUs and 2 GPU WUs (each GPU: 1CPU + 1 ATI GPU)

Note: No screensaver.

WCG Username: AnnCore
CPU Model/Speed: QX9770
GPU 0 Model: ATI Radeon HD 5850
GPU 0 Time per WU: 14:25
GPU 0 Clock Speed: <If overclocked>
GPU 0 GPU Result Name: <optional>

GPU 1 Model: ATI Radeon HD 5850
GPU 1 Time per WU: 13:57
GPU 1 Clock Speed: <If overclocked>
GPU 1 GPU Result Name: <optional>

For all times per WU I took the average of 5 consecutive results.

I haven't tried 4 WUs with screensaver on.

Bottom line, 4 WUs in about 14:11 > 4 WUs in about 17:20 (2 x 2 WUs in about 8:40).

Edit:

4 GPU WUs with screensaver (1 CPU + 1 ATI GPU)

4 WUs every 14:05. (the average of 20 consecutive results)

Not sure what happens with the screensaver on with only 2 GPU WUs...

My son's rig:

WCG Username: AnnCore
CPU Model/Speed: AMD Phenom II X6 1090T
GPU 0 Model: ATI Radeon HD 5850
GPU 0 Time per WU: 8:02 (the average of 10 results)
GPU 0 Clock Speed: <If overclocked>
GPU 0 GPU Result Name: <optional>

[topry]

Thanks for doing this, I've been researching the same info. Following are my stats:

CPU Model/Speed: i7 960/3.2
GPU 0 Model: GTX 660 'SuperClocked' model
GPU 0 Time per WU: 11-14 (1CPU + 1GPU)

GPU 1 Model: GTX 550Ti
GPU 1 Time per WU: 16-18 (1CPU + 1GPU)

Running 2WU at .5 + .5 gives a much broader range, with an average of 30% increase for each card.

Attempts to run 3WU with latest nVidia drivers resulted in driver crash within 15 minutes.