Test to Compare Nvidia Cards for CS5
 

Can someone think of a test using XDCAM EX 1080/30p clips with CS5 using different nvidia cards modified to work with MPE acceleration.

I am going to microcenter to pick up a 9800GT 1GB (112 cores) and a GTX 275 896MB (240 cores).

I want to see what the difference is. And depending on whether my PC Power & Cooling 750w can handle a GTX 470, I might pick one up tomorrow to test as well (I have 10 drives, i7 920 & 2 raid controllers).

I hope to start test tonight.

Sounds like a plan Steve.

Ideally what would be AWESOME would be to put together a package / project with 20-30 second clips of different codecs - xdcam, hdv, avchd, r3d - have timelines for each, and then just progressively stack effects on every 20 seconds, add layers pip's etc.

That way there would be one controlled reference project EVERYONE can download and report playback performance AND render times for. That way we would all have a great idea of what GFX card / system combo is right for their needs.

Paul, I like your idea but the problem is that everyone's PCs are different.

I should mention that your post about the # of cores is what sparked my idea to test.

I am also ordering a FX 3800 along with a HP Z800; thus, I can include the FX 3800 in my testing once I get it.

I will get some avchd footage from my Sony XR500V tomorrow to include in the testing.

Here are my PC specs for testing:
Fresh install of Win 7 x64 Ultimate, CS5 Master Collection
150GB Velociraptor for OS & Apps
i7 920 (stock clock), 12GB ram
Areca 1680ix with 4 74GB Raptors in Raid 0 for video files
1 - 500GB Seagate 7200.11 for Media Cache
ASUS P6T

Here are the 2 cards I picked up to begin testing:
$220 - GTX 275 Micro Center - BFG Technologies GeForce GTX 275 OC 896MB GDDR3 PCIe 2.0 Graphics Card
$100 - 9800GT EE Micro Center - PNY GeForce 9800 GT EE 1024MB GDDR3 PCIe 2.0 Graphics Card

WOW!!!!!!!!!!

So far, the GTX 275 has handled everything I have thrown at it - 3 XDCAM EX layers, 1 black layer with 8 point garbage matte made into a vignette with 50% opacity and overlay mode, and a title. All 3 video layers have these effects: edge feather, RGB curves, scaled to various sizes. A gaussian blur is on one layer, sharpen is on another, levels is on one, one is keyframed to move horizontal over 10 seconds. I am probably forgetting a couple effects since I used so many.

Using CS4, I recently edited a 1080p clip scaled to a 720p sequence and with just RGB curves, my i7 CPU was almost maxed out. Adding a another layer as a vignette was too much for the i7.

Yes it will make isolating different card performance harder but dont forget that when actually rendering out sequences the CPU again becomes the important factor so overall by collecting enough data we would be able to get a good idea of both what different cpu's do for rendering and different GPU's do for playback. SO an overall end to end real world workflow of sorts?

Anyway I like what your reporting back on the 275 - did you want any footage of other cameras? Ive got some stock red footage and some stuff from my t2i so let me know.

Damn, I thought people here would be all over this. So many people keep asking "does this xx card work with MPE" or "how well does this card work".

Well, here is what I have found so far using a GTX 275:

4 avchd layers all scaled to 50% and no effects. adding 1 effect causes it to bog down.

3 avchd layers, all scaled to various amounts; the same 8 effects on each layer and I can add more effects without losing realtime playback.

All effects used are hardware accelerated.

I'd like to see the difference between the GTX 285/GTX 295 and the newer GTX 470/480.

I'm thinking about buying a GTX 470 but not sure its quicker than the 285/295 for Pr use.

The GTX 295 is essentially 2 GTX 275s in SLI but Adobe does not utilize more than one GPU; so, I don't see the 295 providing any benefit over a single 275.

There is a video from NAB showing the GTX 285 as capable of handling 3 layers, FX 3800 = 5 layers and FX 4800 = 7 layers. However, the 285 has more cores than both the FX 3800 & FX 4800, 240 vs 192.

So, I don't know what Adobe is doing to 'limit' the cards based on their price.

Back to testing: I exported 2 mins to H264 / 2 pass / 20Mb average, 28Mb max
It took 8mins 40s.

I know from experience using CS4, the CS5 export time is far quicker.

Good work Steve - I think its very apparent that adobe has indeed locked the number of tracks for the none Quadra cards to 3. The real test now will be to see using the 3 track limit, is there any difference between a 250,260,275?

Indeed but it will be a hard one to test as I doubt anyone has the same setup and all these GPUs

Well Steve said he has both the 275 and the 9800 - the 9800 has as many CUDA cores as the 250 both of which is half the core count of the 275. So by simply comparing those two on the same system we will get a good idea of how much MPE is using the CUDA cores and how many you really need. Its a start.

Some bad news, I had to return both cards today thanks to an idiot who screwed up my paycheck for last month's invoice; thus, I must wait two more weeks for some income. And it doesn't help that half of my invoice was reimbursements for purchases and rentals. Today, Best Buy & Microcenter became my revenue source from all the returns.

Anyways....I bet the 9800GT is enough for 3 layers and accelerated effects. Remember that playback of the original video is still done by the CPU.

However, I bet the faster the GPU, the faster the export times are.

Exporting the same 2min of video takes 46 mins in CS4.

I suspect that there will be no difference between any of the Geforce cards at all. All of the GTX are faster than Quadro (in their specs), but obviously Nvidia has either paid or otherwise made Adobe agree to NOT let the consumer level cards overhaul the Quadros in any aspect, so the slowest Quadro will be the common denominator here. I hope Im wrong, but how else do you explain the lack of support for any cards exept the 285 (is that the one Nvidia had overstocked I wonder?)

Thats bad news Steve, hope you get it all sorted out. Amazing speed boost in the render by the way

Yes Nik my conspiracy theory is that Adobe made it far too easy for us to find a work around and make any cuda enabled card to work. That way we get used to the power and speed but before too long hit the wall when we need more than 3 layers. Then it seems the only option will be to fork out for a Quadro

Unless some friendly hacker finds out what Adobe did in their code and writes a workaround. I know their used to be ways to both softmod and hardmod geforce cards to flash the bios and make them appear as Quadro's

Guess we wait and see...

Steve,

Now that we are back on track about testing, I have two suggestions:

1. Despite all it's current limitations, the PPBM4 home page benchmark can give you a good basis for comparison. It is only AVI material, but it gives you a comparison and since you will be testing with different nVidia cards on otherwise identical systems, the results may be telling about the impact of various video cards. (My results with MPE GTX-480 went down to 14.8 seconds)

2. If you send me a mail how to get in touch with you, I can send you a preliminary ppbm5 benchmark project / timeline with material comprising AVCHD, HDV and XDCAM in 29.97 and 25 fps for you to download. I'll contact you on your mail address.

EDIT: THIS IS NOT HACKING! It is simply opening a txt document in notepad and typing in the name of your video card. If Adobe really didn't want people to do this, then they would have made it very difficult to do so. Heck, I wonder if an Adobe employee is the one who leaked this info.

Actually, this is a great discussion that I would like to continue if a mod could cut out these last several posts and create a new thread about Adobe Stability.

Here is what I see:

Avid: releases features Premiere already has but Avid MC users celebrate as though Avid is sooo great for giving them these features. Avid users treat their workstations with a lot of respect and don't install a bunch of crap software.

FCP: Apple adds a few features Premiere already has but then slows down the software, especially Compressor and FCP users flock to the upgrade. FCP users are inherently limited to the amount of crap software that can be installed since not much is made for OSX.

Adobe: innovates and adds NEW features far before anyone else does. Adobe does try to be everything to everyone, which does hurt its reputation.

Not about hacking? Ok.

At request, 12 posts copied or moved to:
http://www.dvinfo.net/forum/adobe-cr...s5-thread.html

Thanks Pete for splitting the posts.

Not Hacking - I am referring to the process of getting a nvidia card to work with MPE hardware acceleration. Heck, it is far harder to overclock a CPU than it is to make a nvidia card work with MPE. And I wouldn't classify overclocking as 'hacking'.

Gosh, I need to get one of the higher-end NVIDIA cards. My system's current HD 4850 has been falling a little behind now...

(To be specific, my system got only slightly better overall performance with CS5 than it did with CS4...)

I use the http://ppbm4.com/Benchmark.html so see the improvements between CS4 and CS5. I have seen these test results on my own system:

1. CS4 4.2.1 with HD 4870: 38.0 seconds

2. CS5 with HD 4870: 23.3 seconds

3. CS5 with GTX-480: 14.8 seconds

This was on the exact same hardware, apart from the video card. One may argue the validity of the test, but with over 100 systems tested it gives a good indication of performance. Notice that the 3-rd results have not yet been published.

I think that the improvement of CS5 is great and with hardware MPE even stunning.

I can't wait to test my new system and compare it to my very slow AMD chip.

Will be able to post the results Wednesday.

Yeah, I suspect the utility of PPBM4 may be a bit less for CS5 than previously. I'm hoping that Bill G is working on a PPBM5 that accounts for such changes as GPU acceleration with its near absence of need for rendering of preview files to disk in most cases, and the optimized decoding of compressed HD source files in the new Intel processors.

Nevertheless, I'm sure it is still a rough approximation and I finally got around to running PPBM4 yesterday on my new system (980X, 12GB, Quadro FX4800, RAID0 x 4). Although I don't have a monster 12-disk RAID like Harm's Beast, mine still turned in 19.6 seconds at default and 18 seconds while overclocked to 4.133GHz.

So the point is this: as you guys go forward with your GPU testing, PPBM4 will allow you to use the same yard stick for each config. That's good. On the other hand, it is somewhat a render-to-disk biased test so may not evenly reflect GPU benefits.

Harm, I did send both those results to Bill yesterday via the email listed in the PPBM4 web site. Actually, thinking about this a little, it might be nice if PPBM5 (assuming there will be one) split out subsystem performance in addition to one composite number. Then folks would know where to put their money for system upgrades.

Pete,

Bill and I are working on it and trying to use actual footage from AVCHD, XDCAM, HDC, DSLR and maybe RED at various frame rates. And use encoding to H.264 and MPEG2-DVD. Also a modified AVI test is considered for disk intensive testing.

The changes in CS5 require a new approach to testing the various system components, but we are thinking along these lines:

1. a disk intensive test, that measures I/O efficiency
2. a CPU intensive test, that measures CPU performance
3. a GPU/CPU intensive test, that measures the impact of MPE
4. a generic video workflow, that reflects common approaches and is a combination of the earlier 3 tests.

Sounds like a LOT of work!

Keep in mind that the GPU cores come on line in MPE when you use effects...not for decode.

Some have been confused because they can lay out 6 tracks of AVCHD on their Nehalem when the card they have is a GTX 285 (which has been limited to three layers as opposed to the Quadros...), but what you can decode for playback is CPU-dependent. You won't get a red line on a big CPU until you reach the computer's ceiling for decode streams...drive speed likely a part of this as well. With the GTX 285, you'll red line the timeline if you have more than three layers -of effects- for the GPU to handle.

A lesser CPU config will make the GPU help look weaker to a user trying to emulate Adobe's demos because the computer won't be able to decode as much video...so you'll get lots of effects on one or two layers to play with vs lots of layers with effects...even with a new, approved Quadro card.

I haven't noticed these nuances covered In Adobe's marketing approach, but perhaps I've simply not seen the correct brochure...

Ack, I discovered that my HD 4850 has only 512MB of total graphics RAM. No wonder why I did not see as much of an improvement in performance as others did. It seems as though I would need a graphics card with at least 1GB of RAM in order to even make anywhere near the full use of the software mode in CS5.

And despite the lackluster performance of my system in the PPBM4 test under CS5, the software-assisted playback turned out to be much faster and smoother than it ever was on CS4. In order to get those nice benchmark results, I needed either a higher-end NVIDIA card with 896MB or more RAM or a faster drive subsystem than my RAID array's current Intel software RAID 0 setup.

Randall,

My score of 38.0 with CS4 was with a HD4870 with 512 MB. With CS5 it dropped to 23.3. With the same 512 MB video card. The further drop to 14.8 was (I think) caused by MPE, not by the 1,536Mb memory on the new video card.

Harm,

How overclocked are you running your machine right now? I was wondering if I could get close to your settings and then run PPBM4 against the GTX285 I have. I have one of the rarer 2GB models. Would love to see the difference between your GPU and mine. I forgot about the EVGA trade up program and finally remembered when I was 96 days out, 6 days past their cutoff, so no Fermi for me. I will be going to CS5 next week after I finish up a big edit I've been working on for the past month. I fear trying to upgrade before I'm done and delivered, so it will wait until then.

With my 4850, I got only a score of about 47 with CS5 versus a score of about 51 with CS4.

And since the test isn't using the drives very much, I might conclude that it could have been my slower drive config than the higher-ranked scorers. And then, maybe it's the fact that I still have only 6GB of RAM in my machine although standard-definition projects should not have been that memory-intensive. And yes, I did try both my old HD 4850 512MB and a newer HD 5770 1GB with the same results (and that my scores varied very widely from one run to another even at the exact same settings). (I put in a 5770 on my main editing machine only because I moved my 4850 to another rig whose older graphics card died out.) With that, I opted not to submit the scores since they were inconclusive.

Even with such lackluster scores, importing/editing/rendering/encoding the videos whose projects originate in CS5 is faster in CS5 than CS4 ever performed in my machine. And that's more important than a benchmark test that isn't quite optimized for CS5. I can't wait for the release of PPBM5!

Looking at the PPBM4 results list, that rig is OC'd to 3.7 GHz. But that rig also has a 12-drive RAID array.

UPDATE: using a PNY 9800GT 1GB (112 cores) and exporting the same 2 minute clip, it took 15m 34s. Something odd I noticed, the CPU usage was only 40-45% during encode whereas it was 75-85% with the GTX 275.

It certainly appears that the number of CUDA cores plays a significant role in rendering/exporting video.

I found out why I've been getting such lackluster scores in CS5 relative to the other CS5 systems in the PPBM4 list:

It's my RAID setup. It is running off of the exact same SATA controller as my non-RAID system drive - and it is using the same Intel RAID driver for both RAID and non-RAID drives connected to the same controller. Most of the systems that achieve higher scores are using a separate controller for their RAID arrays from the one used for their system drives. Unfortunately, my motherboard does not have a secondary RAID controller for any internal SATA hard drives - and the only secondary RAID-capable controller on my mobo is a Marvell eSATA controller for external SATA drives.

Randall,

Are you editing off a RAID 0? If so, are you creating it via the RAID controller or are you creating a software RAID 0 in windows? If you haven't tried creating the software RAID, give it a try. I have built plenty of machines utilizing only the Intel onboard RAID controllers and have had no issues with performance.

Randall,

Looking at your results, I guess you have much more processes running than is really necessary. Your MPEG encoding times are far worse than I would have expected at the clock speed you are running.

Check with Process Explorer that you have not more than around 40 processes running before starting PR or AME. Check that you have compression and indexing turned off on all your disks, that sidebar is not started and that your services are in line with Windows 7 Service Configurations by Black Viper

Yes, I am editing off of a RAID 0. And my current RAID 0 array was created via Intel's Rapid Storage utility.

Even when I got the process count down to 34 processes and compression and indexing turned off and the configuration set to Black Viper's "Safe" configuration (which I used for now), I still get the same PPBM4 score I had been getting with all of the default processes running.

Perhaps this may be because I was/am using an ATi card (an HD 5770, at that)? (In general, the consumer ATi cards do not perform as well in OpenGL mode as NVIDIA's consumer cards do.) I am suspecting this because the HD 5xxx series - at least with the drivers that are currently available for them - perform rather poorly on some of the critical 2D tasks compared to the older cards or the NVIDIA cards.

Or maybe it's the fact that I am also using the same editing system for other uses, such as gaming or as a media center? The software used in such systems also adds resource-hogging processes (much of which can't be disabled or uninstalled without screwing up those programs).

Even after this lackluster MPEG encoding performance with standard-definition material, I would have to say that transcoding 1920x1080/59.94i Cineform AVI files to Blu-Ray-compatible H.264 AVC is about 20 percent faster in CS5 than it was in CS4 or Vegas.

Randall,

Not trying to be rude, but this is totally off-topic. I started this thread to help people make a decision on which nvidia card they should get.

On topic: so far, it appears that Adobe has not yet optimized Premiere to take advantage of more than 240 cores due to the fact that my PPBM CS5 scores do not differ much from Harm's scores, even though Harm's video card has 4 times the # of cores.

Fair enough. I should have started a new thread about this.