Building a new SuperComputer for Vegas Pro 11!

[Harm Millaard]

Jeff, I haven't the faintest idea for Vegas. The 570 in PR is significantly faster than the 460. Have a look here: MPE Gain - PPBM5

One of the major advantages - in some cases - of the 680 over the 580 is the capability to steer more than two monitors. I forgot to mention that.

For me, with a dual monitor setup and a TV connected but without third party cards like AJA, BM or Matrox, that is a huge advantage. It allows me to bypass the DV deck.

[Gints Klimanis]

Originally Posted by Harm Millaard

Jeff,
Do not be lured by the simple number of CUDA cores in the GTX 680 over the GTX 580. They have completely different architectures and number of CUDA cores are just not comparable.

Thanks, Harm. Is the difference closer to the memory bandwidth difference? There are certain types of math software (lots of operations that depend on few input points and produce few output points) that will benefit from more parallel computing units, but the most common video effects do not fall into that category.

The memory bandwidth limits the computing throughput of the CUDA cores. That is, the performance gain would involve over clocking (OC). I'll try the OC experiment, but so far, my eVGA 560ti decides that it doesn't want to work as GPU coprocessor at 5-7% OC with its built-in air-cooling. I'm debating on whether I should open the water block I bought for the 560ti or invest in water-cooling for the 6x0.

[Gints Klimanis]

Results for overclocking an eVGA 560ti for Sony XDCAM (35 Mbps, 4:2:0) MPEG2 to MP4 15 Mbps VBR :

For CoreClock= 963 (+13%), ShaderClock= 1926 , Memory Clock= 2196 (+7%)
-------------------------------------------------------------------------
Sony AVC mp4= 2:58/3:00 (98%) MainConcept AVC mp4= 1:30/1:35 (95%)

There was no improvement moving to 983/1966/2243 or 943/1886/2462 for MainConcept. 1034/2068/2462 AND 994/1988/2258 were unstable and didn't complete. So, it's not an issue of overclocking the GPU memory.

Overall, I'm not seeing a GPU improvement when paired with an Intel Core i7 980x (no OC). It appears that investing in a better CPU or CPU overclocking results in a significant and linear improvement in render speed when overclocked.

[Kim Olsson]

Iam not going to buy my new computer until july...
So I will study the most recent technologies which is useable to my needs...

For now, I'll think I go for 32GB (2133Mhz quad channel), instead of 64GB (1600MHz).

And Jeff... I belive I will test the new GTX 670, If your right that its all about cuda's, this performce equal to the GTX 680. Actually, the nVidia GeForce GTX 670, with manufacturer like asus, gigabyte, performs better then the original GTX 680 (reference card).

GeForce GTX 680 - |CUDA cores: 1536| Standard memory config: 2048 MB| Relative Graphics Perf. 95x
GeForce GTX 670 - |CUDA cores: 1344| Standard memory config: 2048 MB| Relative Graphics Perf. 84x
GeForce GTX 580 - |CUDA cores: 512| Standard memory config: 1536 MB| Relative Graphics Perf. 75x
GeForce GTX 570 - |CUDA cores: 480| Standard memory config: 1280 MB| Relative Graphics Perf. 68x

SO u see... For $100 - $150 cheaper, you get the new GTX 670 with small differences...

But in mean time, I will investigate further...

And I hope more people jump the wagon, and enlighten us!!

The big queastion is about GPU chip and RAM config. for VP11 =)

[Jeff Harper]

Very interesting Kim, about the 670, please keep us posted!

[Gints Klimanis]

[QUOTE=Kim Olsson;1733803]
GeForce GTX 680 - |CUDA cores: 1536| Standard memory config: 2048 MB| Relative Graphics Perf. 95x
GeForce GTX 670 - |CUDA cores: 1344| Standard memory config: 2048 MB| Relative Graphics Perf. 84x
GeForce GTX 580 - |CUDA cores: 512| Standard memory config: 1536 MB| Relative Graphics Perf. 75x
GeForce GTX 570 - |CUDA cores: 480| Standard memory config: 1280 MB| Relative Graphics Perf. 68x
/QUOTE]

Thanks, Kim. This GFX performance isn't from video editing apps, is it?

[Harm Millaard]

Here is the table with Bill Gehrke's results for PR CS6. I don't know how this translates to Vegas, but at least it shows the benefits of the 680. Adobe Forums: Current sweet spot in GTX-4xx to -6xx graphics card?

[Gints Klimanis]

Thanks, Harm.

[Kim Olsson]

No Gints. Actually its relative to gaming...
Its a measurement so a gamer could compare a graphiccards performance.
Mayby it means something to apps aswell now when they can utilise GPU.

[Peter Siamidis]

Hey guys figured I'd add some input to the fray. I use Vegas Pro 11 daily and do lots of encoding, so I've done many benchmarks on various setups. What I've been using for over a year now is this setup, with approximate prices:

Asus P8p67 motherboard, ~$150
Intel i7-2600k clocked at 4.7ghz, ~$310
Coolermaster V8 cpu cooler, ~$50
16gb 1600mhz ram, ~$80
Corsair AX850 power suppy, ~$160
NVidia 560ti, ~$220

I have Vegas scripts that do much of the work for me, so I run a script and it will add watermarks, fade ins, create directories and render out multiple versions of the videos. Anyways the above setup is relatively cheap, around ~$970 and it's been absolutely rock solid for me. For reference with the 560ti video card Vegas Pro 11 encodes my h264 versions around 4x faster than with just cpu alone, which is really nice. The i7-2600k's are also very overclock friendly, mine runs 4.7ghz all the time for ages now without issue, around 1.400v and ~73C at full load. I've seen people run then higher than that so I might have more breathing room there, but I'm leaving it at 4.7ghz for now.

I was considering switching to an NVidia 670, but because of the more limited design of it's cores (compared to the NVidia 5xx series) in some operations I've read that it may not be as good for cuda use right now. If someone had Vegas Pro 11 benchmarks for that card that would be great to read!

[Gints Klimanis]

Peter wrote "For reference with the 560ti video card Vegas Pro 11 encodes my h264 versions around 4x faster than with just cpu alone, which is really nice. "

This I don't get. Are you using a lot of video FX? I just posted measurements nearly no performance difference when comparing the GPU (560ti) to the CPU (Intel Core i7 980x 3.33 GHz rendering for Sony XDCAM source files with no only video and audio crossfades to 16 Mbps h264 (720p60 or 1080p30). Of course, video preview is off. When I overclock the CPU, there is proportional increase in rendering speed.

[Peter Siamidis]

Originally Posted by Gints Klimanis

This I don't get. Are you using a lot of video FX? I just posted measurements nearly no performance difference when comparing the GPU (560ti) to the CPU (Intel Core i7 980x 3.33 GHz rendering for Sony XDCAM source files with no only video and audio crossfades to 16 Mbps h264 (720p60 or 1080p30). Of course, video preview is off. When I overclock the CPU, there is proportional increase in rendering speed.

I'm not using any fx, just a watermark. I just did a quick render test, the source footage is from a Sony 560v camera and is 1920x1080 60p at 28mbps. I dropped a piece of footage into Vegas 11, and all I added was a watermark to it using the legacy text media generator tool which I've found to render faster than the newer text media generator. I rendered out a 1 minute segment of this video to this spec:

Mainconcept AVC
1920x1080
29.970fps
Two pass encoding
20,000,000 max bps
8,000,000 average bps
Progressive download enabled
2 Reference frames
Deblocking filter checked
High profile

I disabled smart resample. I have cpu and gpu meter gadgets that show how busy each device is. The result:

Cpu only:
5 minutes 30 seconds
Cpu cores bouncing between 92% and 100% use
Gpu at 0%

Cpu+gpu:
1 minute 34 seconds
Cpu cores bouncing between 92% and 100% use
Gpu bouncing between 54% and 71% use

So about 3.5x faster in that case. That's with Vegas Pro 11 build 683 64bit on Windows 7 64bit, and NVidia 296.10 drivers, 4.7ghz i7-2600k and NVidia 560ti (original model).


EDIT: I just did one more test, a bit simpler. Same type source footage and same render target, only difference this time is no watermark so it's the footage alone with no fx's, and I set the encoder to 1 pass. Results to encode 1 minute of footage:

Cpu: 2 minutes 37 seconds
Cpu + gpu: 37 seconds

In that case it's 4.2x faster encode time.

[Randall Leong]

Originally Posted by Peter Siamidis

I'm not using any fx, just a watermark. I just did a quick render test, the source footage is from a Sony 560v camera and is 1920x1080 60p at 28mbps. I dropped a piece of footage into Vegas 11, and all I added was a watermark to it using the legacy text media generator tool which I've found to render faster than the newer text media generator. I rendered out a 1 minute segment of this video to this spec:

Mainconcept AVC
1920x1080
29.970fps
Two pass encoding
20,000,000 max bps
8,000,000 average bps
Progressive download enabled
2 Reference frames
Deblocking filter checked
High profile

I disabled smart resample. I have cpu and gpu meter gadgets that show how busy each device is. The result:

Cpu only:
5 minutes 30 seconds
Cpu cores bouncing between 92% and 100% use
Gpu at 0%

Cpu+gpu:
1 minute 34 seconds
Cpu cores bouncing between 92% and 100% use
Gpu bouncing between 54% and 71% use

So about 3.5x faster in that case. That's with Vegas Pro 11 build 683 64bit on Windows 7 64bit, and NVidia 296.10 drivers, 4.7ghz i7-2600k and NVidia 560ti (original model).


EDIT: I just did one more test, a bit simpler. Same type source footage and same render target, only difference this time is no watermark so it's the footage alone with no fx's, and I set the encoder to 1 pass. Results to encode 1 minute of footage:

Cpu: 2 minutes 37 seconds
Cpu + gpu: 37 seconds

In that case it's 4.2x faster encode time.

If you're seeing much faster performance in GPU mode compared to the CPU-only mode, then most likely your system has a bottleneck somewhere.

[Peter Siamidis]

Originally Posted by Randall Leong

If you're seeing much faster performance in GPU mode compared to the CPU-only mode, then most likely your system has a bottleneck somewhere.

Not sure what you mean, all the cpu cores are pegged near 100% and not throttling due to temperature so they are maxed out both when using the gpu and when not using the gpu during encoding. I also get relatively linear increase in encoding speed when comparing cpu encode time stock, at 4.3ghz and at 4.7ghz. The harddrive has plenty of bandwidth to feed the encoder, even at 2x realtime encoding speed it's only ~56mbps of bandwidth needed and the harddrive can supply far beyond that even on a bad day. My numbers are not unusual, I've read others have the same encode speed increases from using a gpu.

[Gints Klimanis]

Peter, I will try again with the same settings you posted. I don't have any 1080p60 material. Why don't we test with the same video clip? I'll post some Sony XDCAM footage. Since we both have the nVidia 560ti, we have a useful comparison.