Unsupported GPU for CS5

[cts51911]

With all the interest of the "unsupported" GPUs, I thought it was time to start a specific thread.

Please post your questions and experiences.

Hacking is not advised and the unsupported cards are not ready for production use.

You've been warned! 

[cts51911]

Shootnz,

What format do you shoot?  If I remember correctly, it is not AVCHD or a "difficult" codec.

Chris

[shooternz]

mxf (DVCPROHD), avi uncompressed, mov uncompressed, Prores

[metalsaber]

shooternz wrote:

@metalsaber:

How many video layers wil the hacked card run before it tips up or needs to render?

Will do a test tomorrow afternoon/evening.

[shooternz]

That will be interesting. Thanx for doing that and lok forward to the result (even though its kind of academic  to me now)

I think people really want to compare hacked card performance against the Approved cards.

I dont know how that is quantifed all else being equal but the Layer Limits maybe  one way.

For me... I regularly use 6 to 9 Video layers so I am well pleased with my results.

For the Record I am using a Quadro FX3800.

[metalsaber]

shooternz wrote:

That will be interesting. Thanx for doing that and lok forward to the result (even though its kind of academic  to me now)

I think people really want to compare hacked card performance against the Approved cards.

I dont know how that is quantifed all else being equal but the Layer Limits maybe  one way.

For me... I regularly use 6 to 9 Video layers so I am well pleased with my results.

For the Record I am using a Quadro FX3800.

Ok, just a few tests.

1. I put 6 layers of 5D Mark II video into the timeline with 0 filters.  The bar was still yellow and played back smoothly.
2. I then removed the top 3 layers and started with the bottom 3.  I added 3 filters to each (Sharpen, Gaussian Blur, and Brightness/Contrast).  Yellow Bar and Smooth Playback.
3. I then added a clips 4-6 back in with no filters.  Still yellow time line and smooth playback
4. I then removed clips 5-6.  Added on the Brightness/Contrast filter to the 4th video layer.  Clip turns red/pretty much smooth playback with a few little blips.  Nothing major.  Interesting that the bar does not cover the entire 4th video layer length.  This has to do with the bottom 3 layers all being at various lengths with the 3rd layer being the shortest.  So the red bar on layer 4 only extends to the length of layer 3.  Probably since the remaining length of layer 4 is now the 3rd layer of video allowed to have filters.
5. I then added 2 more filters to layer 4 (3 effects total) and the playback is about as smooth as in #4.  Pretty usable imo.
6. If I add a 5th layer of video with 0 effects, the timeline actually goes back to yellow for the length of the 5th layer.  Even with a 50% opacity on layer 5, it still plays back rather smooth.
7. If I add a single effect to layer 5, timeline turns red again.  Playback is a bit more choppy at this point.

Seems like 3 layers of video with effects will keep the timeline yellow.  Adding a 4th video layer with effects turns the timeline red, but still useable.  Odd part was adding a 5th layer with 0 effects on top turned the timeline back to yellow again.

So my guess is that Adobe is limiting the GTX series to 3 layers of video with effects.  Probably has to do with not ticking off those that purchased the Quadro series cards.

[cts51911]

metalsaber,

I tried to duplicate what you did with similar files from my 5dm2.  Pretty similar results

Here is another way to look at it...

I watched all the MPE demos on youtube (below).  This demo particular seems to be the best real world test  4 video streams with effects.  They are doing the demo on a monster machine too.  I'm pretty happy to be close to something similar with my 5dm2 1080p 24 files.  6,8,10 layers....Not part of what I do.  I see the real benefits of GPU when I switch back to SOFTWARE ONLY and I'm back to jerky video just like the video.  With MPE enabled on the GTX 470, editing is very usable with 4 layers+gpu effects (14 total).  Maybe not perfect or as good a Quardo, but good enough for me.  The best part, preview playback is still acceptable even if the line is red with 4 layers in this case.

http://www.youtube.com/watch?v=xso6CGdsl2c

Oh...have yet to see the rendering errors either.

So far, the hack has been good enough for me.

[shooternz]

Thanx for doing the test.

Seems to show the hacks dont make Quadros out of that series of cards.

[chupacabracobra]

shooternz wrote:

Thanx for doing the test.

Seems to show the hacks dont make Quadros out of that series of cards.

if you don't use tons of layers and keep it to three with effects, then I would bet a Fermi would do even better than the Quadros out there (or at least any non G100-type Quadros). G100 CUDA blows away G80 CUDA. Granted some of the increase might need a different programming style, but even same code, as is, should run faster on G100 than ANY G80 chip quadro or not (and let us be honest a Quadro is just a consumer card with, for the most part, just some driver features enabled, such as more more streams allowed; they hardly go to the expense to design a different architecure for Quadro, they need to huge sales of gaming cards to subsidize it all) and I would think a consumer card of same level as a particular QUadro would also do every bit as well so long as you don't use too many layers.

[shooternz]

if you don't use tons of layers and keep it to three with effects, then I would bet a Fermi would do even better than the Quadros out there

If that suits you, that is fine... but I dont need to limit my self in this way which is brilliant...because my workflow uses tons of layers and multiple effects.

I think Chuck Mcs report above also indicates the reason why the supported cards are beneficial.

At the end of the day... editors will buy what they can afford or justify  and then get on with EDITING movies.

(The only reason I persist in this thread is that I think valuable information is coming out for system builders and users of CS5)

eg..What works and  how well it works ...

I will add to the infobase .... I do not have a RAID system.

[Chuck_A__McIntyre]

I wonder if I got a freak card from BFG, or I'm not understanding the 285 GTX layer limitations.  I have stacked 20 layers of Panasonic 1080i P2 media on the timeline, all with varying opacities around 50%, + and -.  I have picture in pictures spinning, a multitude of blending modes, gaussian blur and color corrected clips.  There is a yellow render line above all 20 layers.  Playback at full resolution is smooth as can be without any stuttering. What's up with that? I thought only the high-end Quadro's were capable of that.

Also MP2 Export is around 9 times quicker than on our other machine.

[cts51911]

Chuck A. McIntyre wrote:

I wonder if I got a freak card from BFG, or I'm not understanding the 285 GTX layer limitations.  I have stacked 20 layers of Panasonic 1080i P2 media on the timeline, all with varying opacities around 50%, + and -.  I have picture in pictures spinning, a multitude of blending modes, gaussian blur and color corrected clips.  There is a yellow render line above all 20 layers.  Playback at full resolution is smooth as can be without any stuttering. What's up with that? I thought only the high-end Quadro's were capable of that.

Also MP2 Export is around 9 times quicker than on our other machine.

Chuck,

You HAVE to post a video.

Nobody is going to believe you.

If what you are saying is actually true, then people are going riot.

[Chuck_A__McIntyre]

I restarted my computer twice to see if some limiting setting failed to kick in.  I did shoot a video of the 20 layer timeline and the playback window.  I will post it later tonight after I get home.

.

[cts51911]

Chuck,

What camera did you shoot the P2 files come from?

Thanks, Chris

[Chuck_A__McIntyre]

Chris:

We used a Panasonic HVX200P.  The camera shoots in rectangular pixels and computers (without Premiere reading the pixel aspect ratio in the file) only count the number of pixels contained in the file.  The video from the camera has an actual pixel dimension of 1280x1080 with an aspect ratio of 1.5.  So this horizontal "stretch" information is contained in the P2 file and Premiere stretches the video horizontally across a 1920 pixel computer pixel "grid" (1280 x 1.5 = 1920).  Finished, edited P2 videos shot with this camera are exported (not P2 but h.264 for example) in delivery format with an actual square pixel dimension of 1920x1080.

I shot this video quickly before I left work:

Video of 20 Layers playing back smooth, full res, yellow render line!... BFG 285 GTX

The BFG card is a honker.  This is why I always build computers with full tower cases.  Some of the wiring was in the way, but I just stretched it out of the way and hoped that I didn't unplug anything in the process.  This card  comes slightly overclocked and has 1GB of memory.

[cts51911]

Hi Chuck,

First, thanks for taking the time to do the video and sharing.

Can you briefly describe your computer setup?

driver version for gtx 285

cpu

HD configuration

memory

Full resolution preview?

Whatever it is that you are doing, I'm sure others want to know how you are doing it.

There is so much going on in your video but it all looks smooth.

Chris

[Chuck_A__McIntyre]

Chris:

Here is a link to where I posted the entire hardware specs for the two computers I built for editing at work:

2 i7 920 Computers

Chris asked:

driver version for gtx 285 - The one posted yesterday on the Nvidia site

cpu - i7 920 not oc

HD configuration - system drive Raptor 150 - no Raid

memory - 12GB Crucial 1600 - 2GB sticks

Full resolution preview? - Yes

Before I pulled the old video card out of the system, I uninstalled the Nvidia drivers, restarted, then shut down.  With the new card installed, I tried it as is to see if Premiere CS5 would see the card.  It did.  I then went to the Nvidia site and used their automatic driver detection browser software, downloaded and installed the latest driver.

Here is a scan of the box our 285 GTX came in:

[metalsaber]

Others that have a 285, are you able to do more than 3 layers with effects?  I'm wondering if the text file "hack" is only part of the MPE performance. There could still be other behind the scenes code that is allowing more layers.

If not, you are one lucky man.

[cts51911]

Anyone have any theories why Chucks typical system specs are producing over the top results?

Is it the file format he is using? (panasonic p2)

Is it the way Chuck setup the project?

GPU aside, his system is pushing 20 files around smoothly.  It has been said by a few here in the know that the CPU still decodes alone (without GPU) still.

I don't think luck or the GTX 285 has anything to do with it.  There has to be a rational explanation.

[Chuck_A__McIntyre]

We have an identical system to the one that is getting 20 layers etc.  I bought two of the BFG cards from Newegg.com.  I didn't have time last night to install the second 285 GTX in our second system.  This morning I installed the card following the steps I used last night for our other system.  The second system performs like the first... 20 layers with varying opacity, blending modes, gaussian blur, color correction, spinning picture in picture, full resolution, yellow render line perfectly smooth playback.

Must be the card.

[chupacabracobra]

Chuck A. McIntyre wrote:

We have an identical system to the one that is getting 20 layers etc.  I bought two of the BFG cards from Newegg.com.  I didn't have time last night to install the second 285 GTX in our second system.  This morning I installed the card following the steps I used last night for our other system.  The second system performs like the first... 20 layers with varying opacity, blending modes, gaussian blur, color correction, spinning picture in picture, full resolution, yellow render line perfectly smooth playback.

Must be the card.

I don't really know anything at all about the P2 format, but perhaps it is an exceptionally easy format to decode??

I know that h.264 (used in AVCHD and other wrappings) is a really tough one that kills CPUs. It is FAR tougher than to decode AVCHD than MPEG2 or VC1.

[Chuck_A__McIntyre]

"I know that h.264 (used in AVCHD and other wrappings) is a really tough one that kills CPUs. It is FAR tougher than to decode AVCHD than MPEG2 or VC1."

I think for that reason, those of us who use the Panasonic P2 cameras (ours is the HVX200P) are very glad to be producing in that format.  Working in 1080i, I have taken a full res (1920 x 1080) still export from the timeline.  You open the file in PhotoShop, view it at 100% and the image is very sharp.

Our older system without MPE takes about 40 minutes to export a 3 minute uncompressed AVI 1920x1080 at the maximum settings.  I can export the same clip at the same settings in just over 7 minutes on our newer system that has the i7 920, 285 GTX with MPE enabled.

Not long ago, it just took too long to experiment with certain effects.  Today I have been trying out a clip in Premiere CS5 dynamically linked in After Effects.  I have been playing with a combination of Posterize and Cartoon.  Amazing quick results.  I do have to render, but it renders fast.

[Wil Renczes]

So yes, P2 is a really good codec in terms of performance.  This means your CPU load is a lot lighter than with other HD material.  The 3 layer cap means that the hardware will only be called to do image processing on the first 3 layers; however, keep in mind that, depending on the effects or complexity, with a beefy system, you might get away with realtime what's left.  And, I haven't tried this (I think Dennis suggested this somewhere), but you might be able to fake it out by pushing layers into a nested sequence, so nested layer + 2 more layers might equal realtime.

That being said, I have no explanation for 20 layers.  I think we're missing something here.  Do you have any layers in that stack that have nothing on them (or specifically, nothing transparent/moved so that it essentially optimizes out layers beneath it because you have a blocking layer that's completely opaque)?

[cts51911]

This works on the GTX 470 I have, i7 930, 12gb

I converted to 5d2 1080p/24 files to Dvcprohd 1080i/60

Imported the 6 layers of clips into a Dvcprohd 1080i/60

All 6 layers - 45% opacity, and a mixture of blur, fast color correction, noise and b/w on the lines - Yellow line still

I'm seeing about 60% utilization on both CPU/GPU

Playback is smooth.

Note that line goes red if I add too much GPU effects even if the playback is still smooth.

Chuck:  Is this the project type you used?   Dvcprohd 1080i/60?

[Craisans]

So hypothetically, AVCHD users could also convert to P2 DVCPRO and experience similar playback?

I'm assuming it's a slow process, but there is a free app from panasonic to convert them available
https://eww.pavc.panasonic.co.jp/pro-av/support/desk/e/download.htm#avchd

This is sort of a buzzkill on the advantage of PPro CS5 being able to natively work with AVCHD, but sounds like a nice option if you're an AVCHD user that wants to go nuts.

[Wil Renczes]

I wouldn't exactly call it a buzzkill.  It's the nature of H264 versus codecs with faster decode times.  Let's put it this way:  prior to CS5, AVCHD couldn't playback at all, and people looking for realtime had no choice but to transcode.  (And, conversely, if you look at FCP, well, you have no choice - you're forced to transcode to ProRes.)

Now, CS5 gets you realtime editing multiple streams for AVCHD, and if you really need to stack 20 layers of complex material, well, yeah, you can still transcode if you're looking for realtime, but then I'm curious what has you doing that kind of crazy stacking in Premiere, and wouldn't it make more sense at that point to be doing compositing in AE..?  With that many layers, you probably need to do more complex masking anyway...<shrug>

(Go AE rotobrush! )