“New” graphics card

[John Rofrano]

[Rich Kutnick] “John. It’s also good to know that my choice of a different graphics card worked out in my favor (due to you and several other colleagues here who recommended the ATI 5870).”

I’m glad that card is really working out for you. The Sony team has been working closely with AMD for a while so it makes sense that Vegas Pro performs better with AMD cards. Like I said, I’m really happy with mine too.

~jr

http://www.johnrofrano.com
http://www.vasst.com

[John Rofrano]

[Rich Kutnick] “I found the problem–thank goodness it’s not my AVCHD footage, either!!”

That’s great. It’s really hard to debug FX slow downs but I’m glad you found it and I’m sure NewBlue will be interested in fixing it. They are good that way.

~jr

http://www.johnrofrano.com
http://www.vasst.com

[John Rofrano]

OK, I got my Quadro 4000 back from PNY and did the same tests last night and it does not look good for NVIDIA’s CUDA with Vegas Pro. The Quadro 4000 is in my fairly new Intel Core i7-3930K Sandy Bridge-E 3.2GHz (6-core/12-threads) so the CPU times are better than my 2008 Mac Pro 2 x 2.8 GHz Quad Core Xeon E5462 (8-cores/8-threads) but the GPU renders were actually slower on the NVIDIA Quadro 4000 than the ATI Radeon HD 5870!

Here are the times for the Quadro 4000:------------------------------------------------
Timeline GPU Acceleration OFF (Playback 0.7fps)
------------------------------------------------
Sony AVC - Internet 1080p (CPU Only). . . . 1:10
Sony AVC - Internet 1080p (CUDA). . . . . . 1:08 (0x)
MainConcept AVC - Internet 1080p (CPU). . . 1:17
MainConcept AVC - Internet 1080p (CUDA) . . 1:07 (0.1x)

------------------------------------------------
Timeline GPU Acceleration ON (Playback 2.0fps)
------------------------------------------------
Sony AVC - Internet 1080p (CPU Only). . . . 0:27
Sony AVC - Internet 1080p (CUDA). . . . . . 0:36 (0x)
MainConcept AVC - Internet 1080p (CPU). . . 0:40
MainConcept AVC - Internet 1080p (CUDA) . . 0:33 (0.2x)
I’m not even sure what’s going on here. First I was shocked that the Quadro 4000 could not play the timeline better than 2fps with GPU ON when the Radeon HD was a solid 29.97fps. This is why the render times are so poor. My Radeon HD 5870 rendered MainConcept AVC with GPU twice as fast as the Quadro 4000! (0:15 vs 0:33). I’m beginning to think that I should be using my 2008 Mac Pro as my primary Vegas Pro editing workstation because the Radeon performs better than the Quadro with Vegas Pro.

I also ran the Sony “Red Car” project and the Quadro 4000 could not maintain the 29.97 fps throughout the timeline like the Radeon HD 5870 did. Again quite shocking to me, but I guess it might have more to say about Sony making better use of OpenCL with ATI cards than NVIDIA cards for timeline GPU acceleration.

Here are my “Red Car” timings for both cards in two different computers:
------------------------------------------------
NVIDIA Quadro 4000 w/Core i7-3930K 3.2Ghz 6c/12t
------------------------------------------------
MainConcept AVC Internet 1080-30p . . . . . 1:34 (94 seconds)
XDCAM EX HD 1920x1080-60i 35Mbps . . . . . 1:42 (102 seconds)

------------------------------------------------
ATI Radeon HD 5870 w/Xeon 2.8Ghz 8c/8t
------------------------------------------------
MainConcept AVC Internet 1080-30p . . . . . 0:57 (57 seconds)
XDCAM EX HD 1920x1080-60i 35Mbps . . . . . 1:15 (75 seconds)
As you can see the Radeon outperformed the Quadro in “real-world” Sony render test with a much slower CPU. That is outstanding for the Radeon! It also beat all of the scores in the Sony report where it was faster than the GTX 570 which was the fastest card in the test.

~jr

http://www.johnrofrano.com
http://www.vasst.com

[Rich Kutnick]

So this is looking more like the best 150 bucks I ever spent!! Without this forum I NEVER would have gained such insight! So basically Nvidia with CUDA does not work very well with SVP 12, while ATI/AMD with OpenCL rocks!! Well I now hope that this will help our fellow colleagues worldwide, for it certainly has helped me tremendously! Kudos to John Rofrano and Dave Haynie for bringing this information to the forefront and for John spending so much time and effort testing all of these combinations!! This should speed up our editing/rendering quite a bit!

Rich Kutnick
VIDEO IMPRESSIONS

[Dave Haynie]

[John Rofrano] “OK, I got my Quadro 4000 back from PNY and did the same tests last night and it does not look good for NVIDIA’s CUDA with Vegas Pro. The Quadro 4000 is in my fairly new Intel Core i7-3930K Sandy Bridge-E 3.2GHz (6-core/12-threads) so the CPU times are better than my 2008 Mac Pro 2 x 2.8 GHz Quad Core Xeon E5462 (8-cores/8-threads) but the GPU renders were actually slower on the NVIDIA Quadro 4000 than the ATI Radeon HD 5870! “

Not even slightly a shock to me… I’ve been largely against the “professional” cards for more purposes. Ok, the way the story is told, you have GeForce, totally optimized for gaming, and you have Quadro, which is totally optimized for 3D visualizations: scientific computing, mechanical CAD, etc. On the surface, you get a professional warranty, lots of RAM (which you need for that whole 3D thing), much faster 64-bit OpenGL primitives, and often just one slot used.

But dig in a little, and you’ll usually find that the pro cards are under powered. The Quadro 4000, for example, is based on the same architecture as the GTX4xx series, and it includes only 256 CUDA cores. The GTX580 is based on a newer version of the Fermi architecture with a wider and faster memory bus, 512 CUDA cores, and sure, only 1.5GB RAM vs. 2GB RAM… not important for video.

Then you dig a little deeper and find out some darker secrets. Now sure, part of what you’re paying for with a Quadro or FirePro (AMD pro card) is much better OpenGL support. There’s not much point in spending time on making OpenGL perfect when your main users are gamers. So in part, the professional series fund this effort at nVidia and AMD. And so, while OpenGL does get back-ported to consumer drivers, it’s not a priority. Plus, you don’t have pro driver upgrade cycles based on tweaks for the latest game release. So there is a valid reason to select the pro series.

Though it always struck me that it was weird, since the pro cards are pretty much always using slightly older versions of the same architecture as the standard cards, why they performed better on some 64-bit OpenGL primitives. Maybe the GPU designers took the consumer chips, added some special magic, and simply built a faster card?

Well, no. What actually happens is that the consumer cards have a few of the 64-bit OpenGL primitives intentionally crippled. This became apparent when the GeForce 4xx series came out, and actually performed much worse on some selected CAD visualizations than the GeForce 2xx cards. Further analysis showed that this kind of slowdown was OpenGL specific… performance on similar things in Direct X favored the 4xx, which is what you’d expect. Some folks coded work-arounds and found that using OpenCL or multiple OpenGL operations, the GeForce cards were performing as expected. I don’t think 64-bit OpenGL gets used for video all that often, if ever, but that’s another functional issue for the GeForce cards, even if it’s technically only “a simple matter of software”.

And it doesn’t stop — and I really didn’t mean to just pick on nVidia, but they seem to be asking for it. We’ve more or less documented here that OpenCL performance is kind of bad on nVidia Kepler cards like the GTX6xx series. In fact, GTX5xx will beat GTX6xx on most any Vegas benchmark that’s using OpenCL. Not sure about CUDA. Checking on this a little more, it seems to be pretty suspect. The whole OpenCL/CUDA things, General Purpose GPU use (GPGPU) was kind of a hack early on. But with nVidia and AMD both released new architectures with GPGPU given marquee billing. That’s Kepler, if you’re nVidia. And yet, they GTX6xx series dropped in GPGPU performance.

So a bunch of people are asking out there in internet land: is it just a bad version of OpenCL on the Kepler cards? Maybe they haven’t fully optimized their OpenCL compiler for Kepler? Or is this another instance of intentional crippling?

Also, check this out: https://www.tomshardware.com/reviews/geforce-gtx-titan-opencl-cuda-workstation,3474.html

That’s a pretty good comparison of GPUs, including the ubergaming GeForce Titan and some of the AMDs, run on “pro” apps using OpenCL and CUDA/OpenGL. The nVidia cards start off strong… then there’s a very obvious flip, about the time they go from 3D video to 3D mechanical CAD. That pretty much demonstrates a problem right in the place you’d expect 64-bit OpenCL to be used. Things go where expected for gaming and CUDA, pretty much (no AMD cards in CUDA, not surprisingly). Once you get to OpenCL, the AMD7970 pretty much wins… the AMD HD7xxx series, dubbed the GCN architecture, is their answer to GPU designed for GPGPU computing, as Kepler is. It’s weird… in a few of the benchmarks, the Titan and even GTX680 do pretty well. But then check out the OpenCL image processing… my three year old, $300 HD6970 does seem to be wiping the floor with the current-year $1600 Titan.

So as an engineer, I look at those benchmark graphs. When the nVidias come out on top, it’s usually the case that the gap between cards isn’t that profound, and even when it is, there’s some logic to it. When you see the nVidias not just lose, but lose by significant margins, that suggests to me that “somethin’s broke”. That could be some kind of flaw in the OpenCL performance. But as with the OpenGL, they didn’t break everything, they just targeted a few operations in the consumer OpenCL, and made them run 1/10th normal speed or whatever. Could this be a similar thing here for OpenCL? And why … oh yeah. The Tesla series. nVidia gets over $5,000 for a Tesla K40, which uses the GK110B chip has 2880 CUDA cores @ 745MHz and 12GB memory. The $1600 Titan has a GK110 chip with 2880 CUDA cores @ 889MHz and 6GB memory. I do detect a familiar pattern here… how could the Titan NOT outperform the K40 on OpenCL benchmarks?

-Dave

[John Rofrano]

Dave, very enlightening. It looks like the ATI Radeon HD 5870/6970 is the best card to use with Vegas Pro right now. Would you agree? I believe you said that these cards were pretty evenly matched with the 6970 having a smaller die and power consumption but a slightly slower clock speed compared to the 5870… and the 7xxx series isn’t performing as well with Vegas Pro because of it’s newer architecture. do I have that right?.

~jr

http://www.johnrofrano.com
http://www.vasst.com

[Dave Haynie]

This week, if I had to buy a new GPU card for Vegas, I would buy the HD6970 if I could find one, the HD5870 otherwise. I don’t think there’s a substantial difference between the two. But honestly, if I was about to buy, I’d at least wait to see if Sony has anything new to say at NAB in two weeks.

I did a little more checking into these GPUs. AMD/ATi have pretty consistently done a kind of tick-tock product line update, and also within a product line, they roll out their new stuff for the high end. So the HD5870 architecture, VLIW5, also showed up as the basis for much of the HD6xxx line. These were all 40nm (chip geometry), the lesser HD6xxx cards downsized memory bus width and/or clock speeds to save power and deliver a cheaper chip package.

Then the HD69xx series launched with a new microarchitecture, VLIW4. Applarently, AMD did a detailed analysis of their VLIW5 instruction set and found it did an average of 3.6 instructions per cycle. Not the five that were possible. VLIW4 reduces the parallel ops to four at a time, but adds a few tweaks on top of those. The HD69xx is also in 40nm design rules, but the architecture change cuts out a bunch of transistors, so the power is lower, performance about the same, give or take, depending on the benchmarks. The simpler design also let some things run faster, clock-wise, but really, these seem pretty evenly matched.

So I looked at all the info I had last summer and determined that going to a new GPU card probably wouldn’t give much if anything, and it would cost quite a bit. And more to the point, if I kept my HD6970, I would have a better notion of which direction to upgrade, based on Vegas 13. This was even before I realized that the Main Concept CODEC was crippled for newer cards and more recent nVidia drives may well cripple some CUDA and OpenGL operations.

I have not seen detailed Vegas benchmarks on the 7xxx, and nothering on the R9 series. I would expect these to run faster still for Vegas internal/plug-in GPU acceleration. But you get nothing out of the Main Concept CODEC… it’s going to be 100% CPU. In some of the OpenCL benchmarks I posted, the 7970 was far and away faster than anything else, the first GPU from anyone at 28nm. This was AMD’s first board using their GCN architecture, their answer to nVidia’s Kepler.

The R9 280x is pretty much the same architecture, also 28nm, some potential performance tweaks, once again, this tick-tock approach (pretty common in the semiconductor industry — whole new chip architectures cost lots of money and years to deliver). I would like to get my hands on one and compare, particularly after Vegas 13 ships.

I’d like to think that in two weeks, Sony will be at NAB introducing Vegas 13 with full support for the newer GPUs. So the equation could change pretty soon. Not that a 5870 or 6970 would be any less suitable, just hoping the modern stuff is supported. And maybe Sony will have something to add on the nVidia vs AMD question. Maybe even some info on potential nVidia software imposed performance limitations.

-Dave

[John Rofrano]

Thanks for the that thorough reply. I hadn’t thought about what Sony might do for Vegas Pro 13.0. Good point… we will all have to wait and see.

~jr

http://www.johnrofrano.com
http://www.vasst.com

[Rich Kutnick]

To add smoke to the fire, I have experienced a problem rendering a long-form (1.5 hours) production in SVP 12. With my AMD video card enabled, rendering stops (freezes) approximately 51% through the rendering process, using AVC Blu-ray templates respectively for MainConcept and Sony. During the process it approximates the rendering time at about 3.5 hours. I then went back to my old standby, MainConcept M2V creation, which estimated less than 2 hours’ render time but again froze about half way through. So I disabled my video card and ran the MainConcept M2V Blu-ray template once more, and in an hour and forty-six minutes the rendering completed without a hitch! As in the past, this appears to me to be an SVP 12 anomaly, not one with my PC or peripherals!! The question now begs–has SVP 13 addressed any of these issues? Furthermore, who has the scoop on SVP 13? Is it worth the upgrade, or as in the past is it almost ALWAYS worth the upgrade? I look forward to everyones’ input on both the rendering and SVP 13 news!!

Rich Kutnick
VIDEO IMPRESSIONS

[Jeff Bauer]

Did Pro 13 address any of the support issues for newer video cards as discussed in this thread or do the recommendations made here stand?

Jeff