PDA

View Full Version : Benchspeed or Bunkmark?



Aselert
03-26-2012, 01:45 PM
It might be interesting to create a benchmark as do other programs (Maxwell Benchwell I remember) to enable comparison of different graphics cards capabilities.
I speak mainly GPU even if one could imagine CPU bench or Hybrid. But personnally I fear that it gives too much disparity ...

It could be a number of passes (1000 by exemple) in minimum time or vice versa (limited to 1 min) on the basis of "Material model" with its presets.

Good idea or not?

Cheers

andy
03-26-2012, 04:31 PM
There use to be a scene of some sort of underwater personal transport device I think.

To have a good benchmark though you'd need a few different scenes and render sizes and record all of the numbers.

I'd say a more simple scene, another with self lit objects and glossies, and another with Tons of data and high resolution textures. Render each at a few different sizes.

david.randle
03-26-2012, 05:54 PM
Download "Bunkspeed Graffiti Benchmark" from the Community under Projects.

Then render with the following criteria:

1920x1080 500passes Inline (mm:ss)
1920x1080 500passes Background (mm:ss)
960x540 realtime FPS at 1000 passes

Example:

"1x Quadro 5000 + 2x Tesla C2050"
2.2 (mm:ss)
2.32 (mm:ss)
8.5 FPS

Cyberwest
08-14-2012, 03:12 PM
Hi,

I'd like to benchmark using Bunkspeed, but I can't find a single sample BIF file anywhere, and the Community Project area you mention here doesn't seem to exist. Any suggestions?

Thanks!

Best,
James

JohnG
08-14-2012, 03:25 PM
This is the project David was referring to..

http://www.bunkspeed.com/asset/viewasset?assetId=463

Cyberwest
08-14-2012, 03:29 PM
Brilliant - thanks! This is for 3D World magazine in the UK, by the way.

lele
08-21-2012, 09:53 AM
Bunkspeed Shot 2012.5.1.2 "Trial" GPU only (NO background) - 500 pass - res. 960*540

ZOTAC GTX 560TI - 2GB - 384 Cuda
174s

ASUS GTX 670 - 2GB -1344 Cuda
125s

andy
08-21-2012, 11:40 AM
Ouch! Really LeLe?

With those number it appears Keppler is running almost 2.5x slower than a Fermi card with comparable numbers. I thought Bunkspeed was seeing a slight increase at least.
Using straight up numbers, your 670 Should be 3.37x faster, but it's only 1.39x faster. Since it has way more cores, and only a slightly slower clock you would think it should breeze through the frames.

I'm hoping it's just not quite optimized yet.

lele
08-21-2012, 02:28 PM
Kepler technology works differently than the Fermi. The amount of core is not comparable between their
Kepler cards have a lot more RAM memory and consume much less and the same prices
In my opinion a good step forward, even when the possibility of improving performance

andy
08-21-2012, 03:25 PM
Sure, I think it's a step forward too, but it's pretty darn difficult to compare. With those numbers, none of the Kepler cards compare to the GTX 580 3GB. Not even close.

So we'll have to either wait for some optimization, or the next generation. Granted, if it's that much more power and cooling efficient, we can also jam more cards into a box easier.

But since games are getting better performance on Keppler, I would guess we just need some iRay optimization. I'm still going to wait on any "upgrades" for now. The slowest card we have in the building is the 470, which is still faster than the 670's. I do stuff that doesn't fit on a card often, but the designers don't unless they try and render a suite of products at once. But we have a render station for that rarity.

lele
12-16-2012, 11:44 AM
Bunkspeed Shot 2012.5.4.8 "Trial" GPU only (NO background) - 500 pass - res. 960*540
(I don't know if this version is compiled with cuda 4.2 or cuda 5)

PNY QUADRO K5000 - 4GB - 1536 Cuda
185s

andy
12-17-2012, 11:05 AM
185s with te K5000 really? That is incredibly slow isn't it? It Should be 4 times faster than the 560, but it's actually sightly slower?
At worst it should be comparable to the 670, but it's slower than that as well.

I'm guessing something else in the system is causing a bottleneck here.

david.randle
12-17-2012, 05:03 PM
Sadly no....the K5000 is about 40% slower than the Quadro 5000 meaning that benchmark is dead on. Kepler performance is lacking so far and we all hope that change is on the way.

andy
12-18-2012, 11:07 AM
But 670 is Kepler too, but it's not nearly that slow. That's why I was thinking it wasn't quite right.

lele
12-18-2012, 01:59 PM
Sadly no....the K5000 is about 40% slower than the Quadro 5000 meaning that benchmark is dead on. Kepler performance is lacking so far and we all hope that change is on the way.

@David
the current version of Bunkspeed is compiled with CUDA 4.2 or 5?
I ask because in Vray-RT (cuda 5) the Quadro K5000 is 2,5 times faster than the Quadro 5000

david.randle
12-18-2012, 06:04 PM
the version of iray we currently use is compiled against CUDA 4.2. I imagine a version that is compiled against CUDA 5 is right around the corner. Thanks for the info

jakes
12-26-2012, 09:44 PM
I am lucky to have access to some great workstations during normal business hours and we are currently getting some brand new dual 690 boxes within the next two weeks. BUT...every once and awhile I need to run something at home, usually as soon as the kids are in bed. For these occasions I have the Po' Boy Box. In other words, what's the biggest bang for the buck I can piece together out of an odd mix of 4 year old guts and some brand new pieces. This thing has to serve as CAD workstation for Proe and Alias, Video editing for school sports, games of all kinds, and all other kinds of abuse that a proper box shouldn't be exposed to :)
Right now this consists of:
Asus P5QL VM Mobo with an Intel Core2 Quad Q8300 - no OC
8 gig DDR2 ram
Sata 2 Barracuda Drives
600W Corsair PSU
For graphics, I was running an EVGA gtx 550ti OC prior to Christmas...

Thankfully Santa brought a brand new EVGA GTX 660ti FTW Sig2 for right around $280. I know what you are all going to say about the gimped 104 chips and how Fermi is so much better. Yes it is for a dedicated machine. I could have gotten a 570 card for around $250 but would have needed to go bigger on the PSU and for the other stuff this rig has to do, the 6 series made more of an impact.

Here are the numbers that count:
GTX 550 ti - Bunkspeed Graffiti benchmark 960x540 500 passes - 305 sec! really bad. RT @ 500 passes 1.66 FPS
GTX 660 ti FTW sig 2 - 960x540 500 passes - 147 sec HUGE IMPROVEMENT RT @ 500 passes 3.42 FPS

Huge bang for the buck on a really outdated system. I'll run the same tests on our 580, dual 590 and upcoming dual 690 boxes soon but in a way it's more satisfying to build a sleeper and get good numbers.

Merry Christmas all,
J

jakes
12-27-2012, 02:04 PM
First reference at work:
Dell workstation
Xeon X5670 @ 2.93GHZ
12 GB ram
GTX 680 2048
Bunkspeed Graffiti benchmark 960x540 500 passes - 109 sec. RT @ 500 passes 4.75 FPS

lele
12-27-2012, 10:56 PM
Bunkspeed Shot 2012.6.0.10 "Trial" GPU only (NO background) - 500 pass - res. 960*540

PNY QUADRO K5000 - 4GB - 1536 Cuda
175s

Asus GTX 660TI DC2 - 2GB - 1344 Cuda
142s

Quadro K5000+GTX660TI
81s

jakes
12-28-2012, 02:45 AM
lele,
What kind of FPS are you getting in RT with the two cards together? That Quadro must really have a slow clock. That card doesn't even have active cooling, correct?
The Asus number doesn't surprise me knowing it could be clocked faster or maybe your machine has PCI 16 x3 vs. mine with PCI 16 x2.

lele
12-28-2012, 08:15 AM
lele,
What kind of FPS are you getting in RT with the two cards together? That Quadro must really have a slow clock. That card doesn't even have active cooling, correct?
The Asus number doesn't surprise me knowing it could be clocked faster or maybe your machine has PCI 16 x3 vs. mine with PCI 16 x2.

I can not remake the test fps because the trial has expired yesterday :(
Quadro k5000 has active cooling is the pci-e 16 x2 (gpu-z)
Bunkspeeed Shot release 2012.6.0.10 is a little faster than the previous ;)

jakes
01-06-2013, 04:41 PM
Here's the benchmark run on a Dell mobile workstation:
Dell Precision M6600
i7 2760qm 2.4ghz 16 gig ram
Quadro 5010m 4 gig with 384 cores FERMI
Grafitti benchmark 960x540 500 passes - 197sec
RT 2.46 fps

A good bit slower than my beater box...

Aselert
01-30-2013, 01:02 PM
With the big help of HamdiR, see the "Benchspeed" sum-up:

Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

3x GTX580, 1536 cores = 30s
2x GTX580, 1024 cores = 42s
2x GTX680, 3072 cores = 63s, 59s
2xGTX570, 960 cores = 61s
GTX470 + GTX670, 1792 cores = 70s
4000+C2075, 704 cores = 72s
GTX 580, 512 cores = 80s
K5000 + GTX660, 2880 cores = 81s
GTX 480, 480 cores = 97s
K5000, 1536 cores = 185s, 175s, 106s (big gap no?)
GTX 680, 1536 cores = 116s, 109s
GTX 670, 1344 cores = 126s, 125s
GTX 570, 480 cores = 120s, 93s (big gap no?)
GTX 470, 448 cores = 129s, 127s
GTX 660, 1344 cores = 142s
GTX 560, 384 cores = 175s
5010m, 384 cores = 197s

Bunkspeed Graffiti Benchmark 4000x2250@500 passes

3x GTX580, 1536 cores = 7:36
2x GTX580, 1024 cores = 11:28
2x GTX680, 3072 cores = 15:57
1x GTX570, 480 cores = 22:57
1x GT480, 480 cores = 23:44

I've mentionned "big gap" sometimes, but may be it come from the evolution of Iray, which is faster now (2.0 < 3.0) as I understood...

charltog
02-27-2013, 04:00 PM
Hello all,

I am running some benchmarks on different graphics configurations with Bunkspeed, and I am using the format suggested by David on the first page of this forum topic:


Download "Bunkspeed Graffiti Benchmark" from the Community under Projects.

Then render with the following criteria:

1920x1080 500passes Inline (mm:ss)
1920x1080 500passes Background (mm:ss)
960x540 realtime FPS at 1000 passes

Example:

"1x Quadro 5000 + 2x Tesla C2050"
2.2 (mm:ss)
2.32 (mm:ss)
8.5 FPS


I am finding that when I load a project, Bunkspeed Graffiti or any other, and run the 1920x1080 500 passes Inline, the first run it has a much slower time than any subsequent runs. Even if I start the render (taking note of the projected "time remaining"), then cancel it after a few seconds, and start it again, I get a much lower projected time. Any number of runs after this seem to have the "fast" time. Likewise, if I run the 1920x1080 500 passes Background first, the first run will get the "slow" time and any runs after that get the "fast" time.

Has anyone else experienced this? What could be the cause? And is there something I can do to ensure that I get consistent, accurate scores for these benchmark runs?

david.randle
02-27-2013, 06:52 PM
Hi charltog - to be honest, you can ignore the inline benchmark. We will be removing that render mode in the future as it is causing lots of complication supporting it.

Just and offline benchmark and realtime FPS are good enough.

krk1
07-14-2013, 10:15 PM
Can anyone tell me how to run this benchmark? I don't see any project files and I don't know how to use the rest of the files.

david.randle
07-15-2013, 06:21 AM
just download the Official Bunkspeed Graffiti Benchmark file from the community under projects and follow the instructions at the beggining of this thread.

These tests have historically only applied to iray or the new (accurate) mode, not (fast) mode yet.

krk1
07-15-2013, 08:57 AM
I have the zip file, but it's not clear to me how to open the project. Here is a screenshot of the contents of the zip file. I don't believe there is information in this thread on how to the most basic thing of getting the benchmark open. Sorry if it's very simple, but I just don't see how. I don't see any .BIF file, for example.

http://i42.tinypic.com/35jlc8x.jpg

david.randle
07-15-2013, 04:54 PM
Well that's a problem. It looks like whatever browser you are using is renaming .bif files to .zip files. Not sure how you would fix that. must be a browser setting.

try just renaming the extension to .bif instead of .zip then open it.

Mar10
08-08-2013, 02:03 PM
Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

Quadro FX 3800 (192 Cuda Cores) 9min = 540s :(

Mar10
08-08-2013, 02:05 PM
Forgot to mention the Quadro FX 3800 benchmark was on PRO 2014.

Has anyone benchmarked the Titan or any of the new 7 series GTX cards?

Jay325
09-10-2013, 03:13 AM
1xGTX780 3gig


1920x1080 500passes= 5:14sec
960x540 realtime FPS at 500 passes=6.45

Is this really sad to anyone else?? Looking at some of these other times on here you would think I was rendering in CPU mode!

Mar10
09-11-2013, 02:20 PM
Jay, any chance you have the overall time for the 960x540 @ 500 passes, so we can directly compare with the other times listed? A 1920x1080 image is 4 times the pixels of the 960x540 benchmark image. I would like to say you can divide your time by 4 but I don't think its quit that linear.

blitz
09-11-2013, 03:48 PM
Jay, any chance you have the overall time for the 960x540 @ 500 passes, so we can directly compare with the other times listed? A 1920x1080 image is 4 times the pixels of the 960x540 benchmark image. I would like to say you can divide your time by 4 but I don't think its quit that linear.

Going with your theory Mar10, a 780 is about as fast as a 580, at more than twice the cost. Very dissappointing. I'm going to have to shop around for some used 580's.

Mar10
09-12-2013, 06:19 PM
I don't know for sure but its entirely possible that is where it lies, just slightly better than a 580. If true, I am quite disappointed as well. My twin 780 system was ordered a few days ago for work. I would try to go with used 580s if I was building a my own personal machine. Hopefully we will have the new systems in a couple weeks and I can report a benchmark for 2x780.

Jay325
09-12-2013, 09:21 PM
I'll do a test for you in the next day or so. I just did the same test with 3x580

3x580 3gig memory

1920x1080 500 passes=1:41sec (slower than others with same cards?)
960x540 500passes= 31sec
RT 1000 passes= 18.9fps

Are these 7xx cards always going to be this slow or are we waiting on Nvidia to unlock some magic??

*I'm going to edit my above times to match the 1k passes and add the 960x540 render time.

jakes
09-26-2013, 07:39 PM
Titan test:
2x Titans - 960x540 500 passes 42sec
1920 x 1080 500 passes 2:41
RT 1000 passes= 14fps

Bunkspeed 2014 Nvidia driver 326.41

Aselert
09-27-2013, 06:52 PM
Still problem with Kepler? :( Titan looks very slow too...

Jay325
09-29-2013, 03:03 PM
That's what I was thinking. 2x titans did 500 passes slower than my single 780?

jakes
10-08-2013, 03:58 PM
Some updated numbers on the two Titans...
Titan test:
2x Titans - 960x540 500 passes 37sec
1920 x 1080 500 passes 2:20
RT 1000 passes= 14.7fps

Bunkspeed 2014 Nvidia driver 331.40 and OC enabled

These cards seem to like more power and the OC. I still think there is something not quite optimized in the drivers as they should kill a single 780.

okokoo
10-11-2013, 10:37 AM
2x Titans - 960x540 500passes 40s
2x Titans - 1920x1080 500passes 2:28

Mar10
10-29-2013, 01:45 PM
2 x GTX 780 - 960x540 500 passes 41sec

Everyone's results are showing the same thing. The 780 isn't any better than an old 580. Our new PC upgrades dropped 500 seconds off the benchmark render for us but still, part of me is disappointed in the 780s. I guess I will be happy with 41 seconds...for now.

Aselert
12-03-2013, 09:20 AM
Update of the BENCHSPEED:

Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

3x GTX 580, 1536 cores = 30s
2x GTX Titan, 5376 cores = 40s, 37s
2x GTX 780, 4608 cores = 41sec
2x GTX 580, 1024 cores = 42s
2x GTX 680, 3072 cores = 63s, 59s
2x GTX 570, 960 cores = 61s
GTX 470 + GTX 670, 1792 cores = 70s
4000 + C 2075, 704 cores = 72s
GTX 580, 512 cores = 80s
K 5000 + GTX 660, 2880 cores = 81s
GTX 480, 480 cores = 97s
K 5000, 1536 cores = 185s, 175s, 106s (big gap no?)
GTX 680, 1536 cores = 116s, 109s
GTX 670, 1344 cores = 126s, 125s
GTX 570, 480 cores = 120s, 93s (big gap no?)
GTX 470, 448 cores = 129s, 127s
GTX 660, 1344 cores = 142s
GTX 560, 384 cores = 175s
5010M, 384 cores = 197s
FX 3800, 192 cores = 540s


Bunkspeed Graffiti Benchmark 4000x2250@500 passes

3x GTX 580, 1536 cores = 7:36
2x GTX 580, 1024 cores = 11:28
2x GTX 680, 3072 cores = 15:57
GTX 570, 480 cores = 22:57
GT 480, 480 cores = 23:44

;)

Mar10
12-30-2013, 03:47 PM
2x GTX 780 - Benchmark 4000x2250 @ 500 passes - 12:14 (734s)

Aselert
01-17-2014, 08:27 AM
Update of the BENCHSPEED:

Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

3x GTX 580, 1536 cores = 30s
2x GTX Titan, 5376 cores = 37s
2x GTX 780, 4608 cores = 41sec
2x GTX 580, 1024 cores = 42s
2x GTX 680, 3072 cores = 59s
2x GTX 570, 960 cores = 61s
GTX 470 + GTX 670, 1792 cores = 70s
4000 + C 2075, 704 cores = 72s
GTX 580, 512 cores = 80s
K 5000 + GTX 660, 2880 cores = 81s
GTX 570, 480 cores = 93s
GTX 480, 480 cores = 97s
K 5000, 1536 cores = 106s
GTX 680, 1536 cores = 109s
GTX 670, 1344 cores = 125s
GTX 470, 448 cores = 127s
GTX 660, 1344 cores = 142s
GTX 560, 384 cores = 175s
5010M, 384 cores = 197s
FX 3800, 192 cores = 540s


Bunkspeed Graffiti Benchmark 4000x2250@500 passes

3x GTX 580, 1536 cores = 7:36
2x GTX 580, 1024 cores = 11:28
2x GTX 780, 4608 cores = 12:14
2x GTX 680, 3072 cores = 15:57
GTX 570, 480 cores = 22:57
GTX 480, 480 cores = 23:44

florsignol
02-04-2014, 10:52 PM
Dell Precision T3600, 32 GB RAM
1x Quadro K6000, 2880 cores:

1920 x 1080 @ 500 passes: 4 min 41 sec (281 sec)
960 x 540 @ 500 passes: 1 min 14 sec (74 sec)

Bunkspeed Pro 2014.0.0.0 (Beta release candidate for 2014.1)
CUDA v. 5050
GPU Only

florsignol
02-04-2014, 10:54 PM
Confirming:

2x GTX 580, 1024 cores = 43 sec

Smileface
02-12-2014, 06:05 AM
Dell Precision 490
1x Quadro 4000 256 cores:

960x540 @ 500 passes 3 min 6 sec (186 sec)
1920 x 1080 @ 500 passes: 11 min 31 sec (691 sec)

Bunkspeed Pro 2012.4.5
GPU Only

sergioelatomico
02-12-2014, 06:05 PM
I7 3770 16gb ram quadro fx 4800 1.5gb+ gtx 580 3gb drivers Nvidia quadro 332.50 windows 8.1 64bits
bunkspeed drive 2014.0

gtx 580 512 cores: 960x540@ 500 passes: 70 sec
quadro fx 4800 192 cores: 960x540@ 500 passes: 4 min 45 sec
I7 3770: 960x540: 7 min 07 sec
I7 3770 + gtx 580 ( mode hybrid) 960x540 @ 500 passes: 62 sec
gtx 580 + quadro fx 4800: 960x540 @ 500 passes: 63 sec

haknslash
03-10-2014, 07:50 PM
Bunkspeed Graffiti Benchmark 960x540 @ 500 passes
3x GTX 780 Ti, 8640 cores = 17s

Bunkspeed Graffiti Benchmark 4000x2250@500 passes
3x GTX 780 Ti, 8640 cores = 6m59s

blitz
03-10-2014, 10:16 PM
Bunkspeed Graffiti Benchmark 960x540 @ 500 passes
3x GTX 780 Ti, 8640 cores = 17s

Bunkspeed Graffiti Benchmark 4000x2250@500 passes
3x GTX 780 Ti, 8640 cores = 6m59s

Wow big difference in speed for low res, not so much in high res. Almost half the time of 3x580's in low, little gain on the 4k res.

Aselert
03-11-2014, 07:12 AM
Indeed! The low res is a joke aha!

haknslash
03-11-2014, 12:49 PM
Yes it is interesting how it scales like that. It is at least 400% increase in performance over my Boxx machine using Quadro 4000, Quadro 600 and a Tesla C2075. Loving the speed increase and it's even faster in other apps!!

Mar10
03-14-2014, 02:42 PM
Very interesting to see the difference between the low & high res. Regardless, it's nice to see a card officially trump the 580 which was released in Nov of 2010. How sad is that.

Aselert
03-14-2014, 03:39 PM
Ahaha, yes my cards are destroyed! But the 580 is still very good ;)

Mar10
03-18-2014, 05:15 PM
Yes, the 580 is still a great card! I was merely pointing out how sad it is that it took more than 3 years to produce a better card. If I was building a system for myself, I would be looking for some 580s.

scampbel
03-20-2014, 01:24 PM
So if you guys are setting up a new system you would use 3x GTX 780 Ti? We are going to be specing out new hardware with 3 floating PRO licenses and Powerboost... Just constantly confused on GPU performance... I thought bunkspeed was able to scale speed based on cuda cores.. clearly this isn't the case... Need to know seriously what cards to get... any advice would be greatly appreciated

Titan, K5 or K6 or GTX 780 or ???

Thanks!
Steve C

haknslash
03-20-2014, 08:40 PM
So if you guys are setting up a new system you would use 3x GTX 780 Ti? We are going to be specing out new hardware with 3 floating PRO licenses and Powerboost... Just constantly confused on GPU performance... I thought bunkspeed was able to scale speed based on cuda cores.. clearly this isn't the case... Need to know seriously what cards to get... any advice would be greatly appreciated

Titan, K5 or K6 or GTX 780 or ???

Thanks!
Steve C

Bunkspeed does scale but I'm not sure how linear. In theory if I have 2 GTX 780's it should be twice as fast as a single GTX 780. I haven't done any tests to see how close to ratio it is and how far it drops off (if any) as you add cards.

Are trying to compare the higher CUDA cores of a Kepler card to the lower CUDA cores of a Fermi card and expecting a scaled performance? You would wish that's the case but the archetures are completely different on the two generations. So 1000 CUDA cores of Kepler does not perform the same as 1000 CUDA cores of Fermi. ;)

I would not spend the money on a Titan unless I needed the VRAM. A 780 Ti is a more powerful card than a Titan is, unless you're talking Black. A Titan Black is basically a 780 Ti + Titan, so you get the performance of the 780 Ti and the VRAM of a Titan. NVidia and their marketing games makes it confusing lol. I also would not waste the money on a high end workstation pro card because those are going to render MUCH MUCH slower than the GeForce cards. I've had both setups and for rendering GeForce is the way to go.

If you can find some 580's and don't mind buying used those cards are still going to be your best bang for the buck in terms of performance per dollar. However if you don't or can't buy used and don't need the 6 GB VRAM I would recommend the 780 Ti.

Hope that helps! It will be interesting to see where Maxwell takes us...

sefsdesign
04-24-2014, 12:08 AM
GTX 780 Ti

Bunkspeed Graffiti benchmark 960x540 500 passes - 63 sec

will be improved render times with next Nvidia driver releases?

robl
04-30-2014, 04:59 AM
I got 3 gtx 580 for a machine at work not too long ago for about $400 each, you can still find those around if you google them. From what I've read on this thread the 580's are still the best not just performance to the dollar but performance in general. I render images at about 2000 x1400 with 2000 passes and they take about 3-5 minutes.

blitz
04-30-2014, 02:53 PM
I got 3 gtx 580 for a machine at work not too long ago for about $400 each, you can still find those around if you google them. From what I've read on this thread the 580's are still the best not just performance to the dollar but performance in general. I render images at about 2000 x1400 with 2000 passes and they take about 3-5 minutes.

How are you managing heat and what powers supply do you use?

robl
05-05-2014, 09:59 PM
They get very hot even though they have 3 fans each, we have a regular house fan blowing into them and they still run around 90 celsius at max load. the psu is an antec 1200 watts. We have been thinking of liquid cooling them with the new nzxt gpu cooler shrouds, they're only about 30 plus 60 per cooler.

haknslash
05-06-2014, 02:31 PM
I got 3 gtx 580 for a machine at work not too long ago for about $400 each, you can still find those around if you google them. From what I've read on this thread the 580's are still the best not just performance to the dollar but performance in general. I render images at about 2000 x1400 with 2000 passes and they take about 3-5 minutes.

The 580 is still a great card but I don't think it's the best card to get these days. If you take a look at the benchmark results, that was true prior to the 780 Ti's or Titan Blacks (less so with them due to overpricing) hitting the market. Still you get a lot of bang for your buck with a 580 but as you know with that comes with a price (a lot of heat and power consumption). This is where the Kepler cards shine. Yes you get less performance "per CUDA" in a sense but you gain lower temps and better power efficiency. So it's a trade off. It has taken Kepler quite a while to get to this point of competiting against the previous architecture pound for pound but these latest Kepler cards have dethroned the 580.

CLSLED
06-11-2014, 12:39 PM
Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

1x EVGA GTX 780 Ti = 61 sec.

(HP Z400 workstation)

amoncur
06-23-2014, 06:18 PM
So if I were going to buy a single card, and needed to maximize the vram for a large scene, it sounds like the 780 or the Titan would be the best I can get without spending thousands of dollars. Is that right?

okokoo
06-24-2014, 06:21 AM
A new 800series will come in Q4. (870/880)

Maxwell architecture (2nd gen)
GPU: GM204
3200 CUDA

amoncur
06-24-2014, 04:41 PM
Darn, I need to buy something this week for a big project.

Mar10
06-24-2014, 05:27 PM
The 780ti is probably your best bet right now for a single card, unless you really need the 6GB VRAM on the Titan.

If your system can handle 2 cards and you don't need the higher 6GB VRAM, you might want consider twin 780's since you can find them for under $500 US.

amoncur
06-24-2014, 11:41 PM
Thanks for the reply! If I were to use two 780s, would I be able to make use of both their vram? In other words, if they both have 3GB of vram, would I be able to fit 6GB of geometry in my scene with both cards installed? Or does BS not consolidate vram?

Mar10
06-25-2014, 02:24 AM
VRAM is not consolidated. You are limited to the card or cards with the lowest VRAM. I believe the scene has to be loaded on each card. So the first decision will be if you really need 6GB or if 3 is enough.

amoncur
06-25-2014, 05:02 AM
It will be a very large scene, so I'll need 6GB. Guess it's back to the Titan, eh?

david.randle
06-25-2014, 07:56 PM
Aaron - We handle about 5 million polygons per gig of VRAM FYI.

amoncur
06-25-2014, 08:29 PM
Thanks, David. Is there a way to measure the number of polygons in SW before opening in BS?

Mar10
06-27-2014, 09:05 PM
Just added a 780ti to our systems.


780+780+780ti (7488 cuda cores)

Bunkspeed Graffiti Benchmark 960x540 Resolution @ 500 passes – 26 sec

Bunkspeed Graffiti Benchmark 4000x2250 Resolution @500 passes – 6:38 (398sec)

Mar10
06-28-2014, 03:52 AM
Thought I would update the list since it hasn't been done in a while.

Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

3x GTX 780ti, 8640 cores = 17s
2x GTX 780 + 708ti, 7488 cores = 26s
3x GTX 580, 1536 cores = 30s
2x GTX Titan, 5376 cores = 37s
2x GTX 780, 4608 cores = 41sec
2x GTX 580, 1024 cores = 42s
2x GTX 680, 3072 cores = 59s
2x GTX 570, 960 cores = 61s
GTX 780ti, 2880 cores = 61s
GTX 780ti, 2880 cores = 63s
GTX 580 + FX 4800, 704 cores = 63s
GTX 470 + GTX 670, 1792 cores = 70s
GTX 580, 512 cores = 70s
4000 + C 2075, 704 cores = 72s
K 6000, 2880 cores = 74s
GTX 580, 512 cores = 80s
K 5000 + GTX 660, 2880 cores = 81s
GTX 570, 480 cores = 93s
GTX 480, 480 cores = 97s
K 5000, 1536 cores = 106s
GTX 680, 1536 cores = 109s
GTX 670, 1344 cores = 125s
GTX 470, 448 cores = 127s
GTX 660, 1344 cores = 142s
GTX 560, 384 cores = 175s
4000, 256 cores = 186s
5010M, 384 cores = 197s
FX 4800, 192 cores = 285s
FX 3800, 192 cores = 540s


Bunkspeed Graffiti Benchmark 4000x2250@500 passes

2x GTX 780 + 708ti, 7488 cores = 6:38
3x GTX 780ti, 8640 cores = 6:59
3x GTX 580, 1536 cores = 7:36
2x GTX 580, 1024 cores = 11:28
2x GTX 780, 4608 cores = 12:14
2x GTX 680, 3072 cores = 15:57
GTX 570, 480 cores = 22:57
GTX 480, 480 cores = 23:44

ralexandre
07-09-2014, 10:17 AM
Hi, I'm recommending a hardware config for a client. Do you have any info on the Titan Z? Or a dual 780Ti config would be better?

Alex_Ji
07-13-2014, 03:38 PM
Quadro 4000 + GTX 680
5:18
4:54
6.56 fps

Alex_Ji
07-13-2014, 05:37 PM
Assuming more cores should result in faster rendering, meaning less time;
Assuming number of cores and rendering time is inverse proportional;
http://i.imgur.com/G3FnWza.png

For GF cards, number of cores * rendering time should be consistent if they ARE inverse proportional. As you can see in the chart below, they scatter mostly around 45000 and 68000 (average of 56000).
http://i.imgur.com/7iz03Ta.png

For GK cards, they scatter mostly between 125000 and 275000 (average of 204000). That 150000 range is a lot comparing to 23000 in Fermi cards.
http://i.imgur.com/XR0NUfQ.png

If you compare 3x580 result 46080 with 2x580 result 43008, they are quite close, and that is why I made the 2 for-mentioned assumptions.

If everyone provides data in the same format, as:

Download "Bunkspeed Graffiti Benchmark" from the Community under Projects.

Then render with the following criteria:

1920x1080 500passes Inline (mm:ss)
1920x1080 500passes Background (mm:ss)
960x540 realtime FPS at 1000 passes

Example:

"1x Quadro 5000 + 2x Tesla C2050"
2.2 (mm:ss)
2.32 (mm:ss)
8.5 FPS
My analysis may come more complete. fps can be added to analysis. If temperature and clock speed are also included, that would be even better.

For now and for what we have so far, my guess is GPU clock speed affects the result(rendering time) as well. Among Fermi cards, close speed is closer to each other comparing to Kepler cards. Typical GTX550Ti is clocked at 900MHz, and GTX580 at 772MHZ, with 128MHz difference, or 16.6%. There is no boost clock. Typical GTX 760 is clocked at 980MHz and boost clocked at 1033 (5.4% gain), while GTX780 at 863MHz and boost at 900MHz (4.3% gain), with 117MHz or 13.6% (base to base), 133MHz or 14.8% (boost to boost), 170MHz or 19.7% (high boost to low base). With GPU boost 2.0 in GK700 series, if you have better cooling, boost clock usually goes beyond advertised. That means the test results with the same cards from different person could be different because of temperature difference causing clock difference.

Why does all this matter?
If you are planning to buy a card just for Bunkspeed rendering, use GF average of 56000 divided by total CUDA cores on that or those GF cards, you get roughly the time needed to render that benchmark. For GK cards, use GK average of 204000.

I have uploaded the excel file I have been working on. Feel free to add your own data.
https://onedrive.live.com/redir?resid=D647C2F1F807AA6A!1727&authkey=!AKtjzRgvJxWUZm4&ithint=file%2c.xlsx

Aselert
07-18-2014, 08:30 AM
Very good analysis Alex Ji!

I'm little bit fan of this stats :P
I will include my 7x GTX580 in the Bunkmark when I have time ;)

Aselert
10-07-2014, 09:01 AM
Quadro 4000 + GTX 680
5:18
4:54
6.56 fps


Really for the 960x540 benchmark? It sound like one of you cards is off, no?

Aselert
12-09-2014, 08:20 AM
UPDATED ;)

Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

3x GTX 780ti, 8640 cores = 17s
2x GTX 780 + 780ti, 7488 cores = 26s
3x GTX 580, 1536 cores = 30s
2x GTX Titan, 5376 cores = 37s
2x GTX 780, 4608 cores = 41sec
2x GTX 580, 1024 cores = 42s
2x GTX 680, 3072 cores = 59s
2x GTX 570, 960 cores = 61s
GTX 780ti, 2880 cores = 62s
GTX 580 + FX 4800, 704 cores = 63s
GTX 470 + GTX 670, 1792 cores = 70s
GTX 580, 512 cores = 70s
4000 + C 2075, 704 cores = 72s
K 6000, 2880 cores = 74s
GTX 780, 2304 cores = 79s
GTX 580, 512 cores = 80s
K 5000 + GTX 660, 2880 cores = 81s
GTX 570, 480 cores = 93s
GTX 480, 480 cores = 97s
K 5000, 1536 cores = 106s
GTX 680, 1536 cores = 109s
GTX 670, 1344 cores = 125s
GTX 470, 448 cores = 127s
GTX 660, 1344 cores = 142s
GTX 560, 384 cores = 175s
4000, 256 cores = 186s
5010M, 384 cores = 197s
FX 4800, 192 cores = 285s
FX 3800, 192 cores = 540s

Bunkspeed Graffiti Benchmark 4000x2250@500 passes

2x GTX 780 + 780ti, 7488 cores = 6:38
3x GTX 780ti, 8640 cores = 6:59
3x GTX 580, 1536 cores = 7:36
2x GTX 580, 1024 cores = 11:28
2x GTX 780, 4608 cores = 12:14
2x GTX 680, 3072 cores = 15:57
GTX 580, 512 cores = 19:44
GTX 780, 2304 cores = 19:57
GTX 570, 480 cores = 22:57
GTX 480, 480 cores = 23:44

TUCID
03-18-2015, 09:54 AM
Just a question. How can it be that my single GTX 580 with 512 cuda cores perform similar to 2X GTX680 with 3072 cores? I thought Bunkspeed was all about number of cuda cores. My findings are based on the graffiti benchmark 960X540 , 500 passes- running in the hybrid mode- with SHOT open during the render. It did the test in 59 sek.

david.randle
06-04-2015, 06:51 PM
not all CUDA cores are the same meaning a Fermi CUDA core (like on your 580) is not equal to a Kepler CUDA core. Comparisons on performance based on CUDA core count are only really relevant when using the same generation card and also the same chip size. It gets a bit complicated in that sense.

Aselert
11-24-2015, 11:45 AM
Bunkspeed Graffiti Benchmark 960x540 @ 500 passes

3x GTX 780ti, 8640 cores = 17s
2x GTX 780 + 780ti, 7488 cores = 26s
3x GTX 580, 1536 cores = 30s
2x GTX Titan, 5376 cores = 37s
2x GTX 780, 4608 cores = 41sec
2x GTX 580, 1024 cores = 42s
2x GTX 680, 3072 cores = 59s
2x GTX 570, 960 cores = 61s
GTX 780ti, 2880 cores = 62s
GTX 580 + FX 4800, 704 cores = 63s
GTX 470 + GTX 670, 1792 cores = 70s
4000 + C 2075, 704 cores = 72s
K 6000, 2880 cores = 74s
GTX 780, 2304 cores = 79s
GTX 580, 512 cores = 80s
K 5000 + GTX 660, 2880 cores = 81s
GTX 570, 480 cores = 93s
GTX 480, 480 cores = 97s
K 5000, 1536 cores = 106s
GTX 680, 1536 cores = 109s
GTX 670, 1344 cores = 125s
GTX 470, 448 cores = 127s
GTX 660, 1344 cores = 142s
GTX 560, 384 cores = 175s
4000, 256 cores = 186s
5010M, 384 cores = 197s
FX 4800, 192 cores = 285s
FX 3800, 192 cores = 540s

Bunkspeed Graffiti Benchmark 4000x2250@500 passes

2x GTX 780 + 780ti, 7488 cores = 6:38
3x GTX 780ti, 8640 cores = 6:59
3x GTX 580, 1536 cores = 7:36
2x GTX 580, 1024 cores = 11:28
2x GTX 780, 4608 cores = 12:14
2x GTX 680, 3072 cores = 15:57
GTX 780ti, 2880 cores = 16:08
GTX 580, 512 cores = 19:44
GTX 780, 2304 cores = 19:57
GTX 570, 480 cores = 22:57
GTX 480, 480 cores = 23:44

blitz
11-24-2015, 10:43 PM
Where are the 980's? What did you update?

david.randle
11-25-2015, 12:02 AM
All Maxwell based cards will be supported by SolidWorks Visualize so benchmarks will surface closer to release.

We are seeing expected and noticeable gains throughout BETA.

Aselert
12-02-2015, 12:40 PM
blitz, I've just updated with my GTX 780 Ti ;)

blitz
12-04-2015, 03:00 AM
blitz, I've just updated with my GTX 780 Ti ;)

Wow much faster than the 3x580's. Tempted to maybe get 780ti's now! ;)

Aselert
01-07-2016, 07:38 AM
Wow much faster than the 3x580's. Tempted to maybe get 780ti's now! ;)

Hummmm no blitz! Only faster than single GTX580 ;)