Updated list, links kindly edited by Sebastian so hopefully they'll work now!


Current GPU statistics ~ Average Run Time to Complete 1 MW v1.02 106.88 credit WU :-

HD 7970 (GPU 1200 MHz, RAM 1550 MHz) .............. 21.3s ... tictoc (Win 7)
R9 280X (GPU 1185 MHz, RAM 1600 MHz) .............. 23s ...... DutchDK
R9 280X (GPU 1125 MHz, RAM 975 MHz) ................ 23.1s ... Arivald Ha'gel
R9 280X (GPU 1180 MHz, RAM 1550 MHz) .............. 23.13s . khrylxtko
R9 280X (GPU 1100 MHz, RAM 1500 MHz) .............. 23.5s ... TennesseeTony
HD 7970 (GPU 1200 MHz, 1550 MHz) ..................... 23.7s ... tictoc (Win 10, also a different PC)
R9 280X (GPU 1100 MHz, RAM 950 MHz) ................ 24.1s ... Arivald Ha'gel
HD 7950 (GPU 1188 MHz, RAM 1250 MHz) .............. 24.1s ... GLeeM
HD 7970 GE (GPU 1000 MHz, RAM u/c 1375 MHz) .... 25s ...... DrPop
R9 280X (GPU 1070 MHz, RAM 1600 MHz) .............. 25s ...... WES
GTX Titan (stock) ............................................... *25.2s ... Sebastian *Time derived from running 8 WUs at once
HD 7950 (GPU 1080 MHz, RAM 1250 MHz) .............. 27s ...... DR4GOON
HD 7950 (GPU 1000 MHz, RAM 1250 MHz) .............. 28.1s ... GLeeM
HD 7950 (GPU 1050 MHz, RAM 1250 MHz) .............. 29s ..... DR4GOON
HD 7950 (GPU 925 MHz, RAM 1250 MHz) ............... 31.2s ... salvorhardin
HD 7950 (GPU 900 MHz) ...................................... 33s ...... Tom*
HD 6950 (bios unlocked to 6970, GPU 1000 MHz) .... 34s ..... khryl
HD 7870 XT Boost (GPU 975 MHz, RAM 1500 MHz) .. 34.1s ... kozicka
HD 7950 (stock) .................................................. 39.8s ... Rudy Toody
HD 6950 (bios unlocked to 6970, GPU 820 MHz) ...... 42s ...... khryl
HD 7950 (stock) .................................................. 42.8s ... wayliff
HD 5970 (GPU 900 MHz, RAM 1200 MHz) ................ 43.7s ... superpower (dual GPU card)
HD 5870 (GPU 850 MHz, RAM u/c 600 MHz) ............ 46.2s ... swiftmallard
HD 5850 (GPU 850 MHz, RAM u/c 500 MHz) ............ 51s ..... Assimilator1
HD 5970 (GPU 735 MHz, RAM 1010) ...................... 52.6s ... superpower (dual GPU card)
HD 5850 (stock) .................................................. 60s ..... Assimilator1
HD 4850 (stock) ................................................. 121s .... TennesseeTony
GTX 980 (GPU 1417 MHz, RAM 6000 MHz) ............. 130s .... Biodoc (Linux)
HD 7850 (GPU 1030 MHz, RAM 1200 MHz) ............. 139s ... usao
GTX 780 Ti (GPU 1084-1111MHz, RAM 6800 MHz) ... 157s ... DR4GOON
HD 4830 (stock) ................................................. 161s .... TLS
GTX 970 (GPU 1114-1253 MHz) ........................... 165s .... Orange Kid
GTX 560 Ti (stock) ............................................. 252s .... zzuupp
R9 m275 (stock) ................................................ 290s .... waffleironhead
GTX 670 (stock) ................................................. 363s .... zzuupp
GTX 690 (stock) ................................................. 394s .... S@NL - JBG (dual GPU card)
GTX 750 Ti (stock) ............................................. 445s ..... TLS
GTX 260 c216 (stock) ......................................... 454s ..... Assimilator1
GT 610 (stock) ................................................. 1827s .... M0CZY

Current CPU statistics ~ Average CPU Time to Complete 1 MW v1.0x 106.88 credit WU :-

Intel i7-5820K (o/c 4.1 GHz) ............................... 2055s ... TennesseeTony
Intel C2Q Q9550 (o/c 3.6 GHz) ............................ 3159s ... Assimilator1
Intel Xeon X5570 (Turboing to 3.2 GHz) ............... 3376s ... DutchDK
Intel Xeon X5650 (o/c 4.1 GHz) ........................... 4130s ... DR4GOON
AMD Phenom 940 (3 GHz) .................................. 4680s ... TLS
Intel i7-4702MQ @2.2GHz, (HT on) ...................... 6245s ... Maxwell
AMD X2 4600 (2.4 GHz) ...................................... 9780s ... TLS

************************************************************************************

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

I need to run 8 WUs in parallel, so i can get 99% GPU usage.

Average time per WU, if it would one be one:

MilkyWay@Home v1.02 (opencl_nvidia)
25.23s (201.82s/8)

Milkyway@Home Separation (Modified Fit) v1.36 (opencl_nvidia_101)
7.58s (60.624s/8)

these numbers cannot be accurate.

for one, my r9 280x @ 1190mhz (the card that probably has the fastest crunching times of milkyway atm)needs 23 seconds for 1 MilkyWay@Home v1.02 and 11 seconds (~5 of which are cpu time) for Milkyway@Home Separation (Modified Fit) v1.36.

secondly, you cannot just divide by 8 because then you ignore the fact that the gpu is not on full load with 1wu/gpu. for example, my card crunches the MilkyWay@Home v1.02s in 37 seconds with 2 parallel workunits, and the modfits ones only take between 13-19 with 4 parallel workunits
so by your reasoning, i could then go ahead and divide 37 by 2 and would get a single workunit speed of 18.5 seconds for the 1.02 app and 4 seconds (median) for the modified fit apps.

what is the load of your gpu when you crunch one work unit only? is this an exclusive nvidia issue? my card is on full load even with one work unit (1.02 app; separation is different), but due to the cpu time it needs toward the end it goes down to 0% for ~ 2 secs which is why i have it run multiple at the same time.

Yea the low GPU load on 1 WU is mainly an issue with just the Titan, AFAIK.

I looked through his results (see his link he posted) & can confirm at least they he is getting (mostly) 194-200s WU times for MW v1.02 app.

Good point about the load drop off at the end of a single WU, as you say, then running multiple WUs gives an unfair advantage due to no down time.
Maybe I shouldn't or relented on my single WU requirements? But then we'll have no even vaguely accurate times for Titans.......

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

Good point about the load drop off at the end of a single WU, as you say, then running multiple WUs gives an unfair advantage due to no down time.

i wouldnt really call it unfair, it is just getting the maximum power out of your card by keeping it on (in my case) 99% load at any given time. this is especially important because the modified fit work units require so much cpu time (on my computer, it's between 5-7 seconds with 4 workunits at once, and during that time the gpu just goes on to the next wu while the cpu finishes the first one)

i believe your 1 work unit requirement is reasonable for a thread like this because crunching more than one at a time messes with the times so seriously that you cannot really get any decent information out of it.

Yea I meant unfair in relation to the benchmarking times, not general usage.
But at least the benchmark WUs used are not the mod fit ones.

I wonder if the downtime on 1 MW v1.02 WU can be estimated so as to add the downtime to the single WU time for the Titans? And thereby give a more realistic figure.....

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

1161039572 854981470 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:16:16 UTC Completed and validated 29.27 5.41 26.74 Milkyway@Home Separation (Modified Fit) v1.36 (opencl_ati_101)
1161039573 854984594 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:16:16 UTC Completed and validated 29.27 5.22 26.74 Milkyway@Home Separation (Modified Fit) v1.36 (opencl_ati_101)
1161039575 854996203 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:17:26 UTC Completed and validated 32.45 5.22 26.74 Milkyway@Home Separation (Modified Fit) v1.36 (opencl_ati_101)
1161038676 853157366 292082 21 Jun 2015, 16:12:24 UTC 21 Jun 2015, 16:16:16 UTC Completed and validated 28.27 5.34 26.74 Milkyway@Home Separation (Modified Fit) v1.36 (opencl_ati_101)
1161037530 855059676 292082 21 Jun 2015, 16:11:12 UTC 21 Jun 2015, 16:16:16 UTC Completed and validated 26.26 5.44 26.74 Milkyway@Home Separation (Modified Fit) v1.36 (opencl_ati_101)

1161039945 855061282 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:16:16 UTC Completed and validated 49.47 2.84 106.88 MilkyWay@Home v1.02 (opencl_amd_ati)
1161039946 855061283 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:19:42 UTC Completed and validated 65.59 2.94 106.88 MilkyWay@Home v1.02 (opencl_amd_ati)
1161039947 855061284 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:18:35 UTC Completed and validated 71.57 2.66 106.88 MilkyWay@Home v1.02 (opencl_amd_ati)
1161039948 855061285 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:17:26 UTC Completed and validated 68.57 2.83 106.88 MilkyWay@Home v1.02 (opencl_amd_ati)
1161039950 855061287 292082 21 Jun 2015, 16:13:54 UTC 21 Jun 2015, 16:19:42 UTC Completed and validated 66.69 2.63 106.88 MilkyWay@Home v1.02 (opencl_amd_ati)

the numbers in bold is the cpu time which is included in the run time, and should represent the exact downtime of the gpu.
the 1.02 ones are btw a good example of how the runtime fluctuates with multiple workunits

Regarding your fluctuating WUs times, was that when running multiple units at the same time?
Btw I'd already considered that for the Titan's time, firstly his time was an average of 8 WUs, 2nd I visually scanned 300 of his valid WU times & approx. 3/4 of the results were done in 194-200s, & so I considered that his 201s time was fair.

That just leaves the downtime that would occur when running 1 WU at a time (which needs to be added to the 'single' WU time).

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

yes it was when running 4 at once. the gap in the numbers becomes bigger when you have your pc crunch separation modfits as well because they require so much cpu time in the end

since he stated that he doesnt get full load with only one work unit i guess theres no way to get an accurate number for 1 1.02 app.

Hmm ,the CPU time for the Titan's WUs are surprisingly high at 30-35s for the 194-200s WUs, see here http://milkyway.cs.rpi.edu/milkyway/results.php?hostid=593293&offset=0&show_names=0&state=4&appid=3

Not sure what to do now!......

Divide the CPU time by 8??

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

depends on how many cores he has and whats his cpu usage settings in boinc i believe

when i set 0.5 gpus, i set cpus to 1 and when i use 0.25gpus i set cpus to 0.5 so dont know..

Hi khryl, the Nvidia GTX Titan Black is a special case in my opinion. I think the Nvidia WUs are not optimized for a Nvidia card with that much double precision power. (Titan Black has 1,5 Terraflop, but normal cards are more aroun 0.1 Tflop to max 0.2 Tflop)

When i just run one WU on the Titan Black the card runs at 12% load (double precision enabled in Nvidia control center) or at 25% load (when double precision is not enabled)

I have a i7 4790k CPU in this computer, and when double precision is disabled the CPU-load is near to 100% when i run 8 WUs at once, and the times are insanely high then.

khryl you can run 4 WUs at once on the 7970 / 280X so you have no downtime. And i think not that this is cheating, since the Milkyway@home project gets more WUs returned. But you should also set the CPU to .25, so you have one core reserved for all 4 Milkyway WUs. But keep in mind that your GPU will run at 100% all the time.

When i take a look at wikipedia and compare the 7970 (at 1GHz) to the Titan Black at stock, then the AMD card is way better than the Nvidia card. 7970 has 1 Tflop and the Titan Black has 1.5 Tflop. But they both have comparable times.

[TA]Assimilator1 you have to edit the list, i have a Titan Black, not a normal Titan. The Titan Black has more shader units and a slightly higher GPU Clock. I am not sure what to think about the high CPU times. I have no idea how they are calculated. And the enabled Hyper Treading might mess up the times as well. But i only run Boinc on 4 cores (3 for CPU stuff and 1 for Milkyway exclusive)

To khryl and [TA]Assimilator1:

I could run some WUs on a GTX 680 or 660Ti, or even on a 580. So we have some GPU and CPU times to compare. One WU on the Titan would also be possible.

One problem about the downtime in Milkyway is, that the GPUs got more powerful over time, but the WUs are still need the same amount of crunching.

I wonder how a Fire Pro W9100 would do. It has 2,5 Terraflop double precision crunching power. 2.5 times more than a 7970.

By the way, can you run Milkyway on Hawaii GPUs?

Yea Hawaii's can run it.

You can run it on your other Nvidia cards if you want, not sure how that'll help though?

Re editing list, can't edit here after 1hr, but will do at AnandTech.

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

I haven't seen a GTX 680 or 660ti on your list. So we would have times for those GPUs. I can imagine, that even a GTX 580 would outcrunch a 680 here in milkyway.

We can compare the GTX 680 to the Titan (both Kepler) so we might get an idea why there are so high CPU-Times on the Titan

I will post the times later here. I can also confirm the HD 4850 and HD 5850 times. The times from both cards are still valid, when i look at the times from my cards.

Ah yea, good point, then throw us your times! :)

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

Here we have some times:

GTX 580: 114,2s CPU 21.3s
GTX 460: 292s CPU 52s
GTX 680: 343s CPU 1,8s
GTX 660ti 383,3s CPU 2,4s

All times are from running one WU on one GPU

The GTX 460 outcrunches the 680. I find that a little surprising. But when i think about other projects, the 4xx and 5xx series were stronger crunchers than the 6xx series, at least until the apps were optimizes. In case of milkyway i thing that the 4xx serie also has more double precision power.

The CPU-Times are also intresting, and i link them to the architecture of the GPU.

The 680 runs at stock clocks. the other cards have slight factory overclocks. I have to check them tomorrow.

Average of at least 5 valid results right?

Yea give us the clock speeds of the o/c ones & I'll add the times, thanks :)

MW is all about DP power.

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

I looked at more then 5 results, and guessed the times. But looked more at the longer times. So the times might be a few seconds longer than average.

Here are the times again with clocks:

GTX 580: 114,2s CPU 21.3s (815 GPU / 1025 Mem) (Stock would be 772/1002) and 90%+ GPU Load

GTX 460: 292s CPU 52s (763 GPU / 950 Mem) (Stock would be 675/800) and 95%+ GPU Load

GTX 680: 343s CPU 1,8s (At Stock speeds 1006 GPU / 1500 Mem, but GPU Boost around 1070 to 1080) and 55% GPU Load

GTX 660ti 383,3s CPU 2,4s (At Stock speeds 915 GPU / 1500Mem, but GPU Boost runs at 1058) and 50%+ GPU Load

You estimated the times with 114.2 & a 383.3s? I'll drop the decimal numbers ;).

Some odd times with your 680, but yea 343s seems about right, & I agree with your 1st 2 cards times as well but the 660Ti time seems off http://milkyway.cs.rpi.edu/milkyway/results.php?hostid=623256&offset=0&show_names=0&state=0&appid=3 , looks more like about 410s.

I've added the times for just your 1st 3 cards for now.

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

the 114.2 and 383.3 had only those times, as good as no variations.

The lastest times from the 660ti are with Vlc running, and trying to run 2 Work Units at once. Still impressive to see so little difference between 1 WU 383s and 2 WUs 410s.