Best Graphics Card for CS5

[Steve Kalle]

Randall is 100% correct on the ram requirement. Someone with a 768MB card learned this early on in the 'hack' thread.

Something else to consider is the warranty. Some have a few years and some have lifetime (PNY & EVGA, I think).

[Harry Lender]

Originally Posted by Jon Larson

Harm--The GPUSniffer app actually says it needs 765MB, so a 768 might work--here's a sample from my laptop:

"CUDA Device # 0 not choosen because 765MB are required, and 111MB are present."


And to the OP, we just purchased this card: Newegg.com - GIGABYTE GV-N460OC-1GI GeForce GTX 460 (Fermi) 1GB 256-bit GDDR5 PCI Express 2.0 x16 HDCP Ready SLI Support Video Card

It has two fans which are inaudible from outside the case, even after 30 mins of CUDA enabled encoding in Premiere.

Hi Everyone:
I just ordered the above Card from Newegg.com. I was impressed with what I read. Plus the fact that it's in my price range. I report back after I install it and put it to use.
Thanks again for all your replies.
Harry

[Randall Leong]

Originally Posted by Steve Kalle

Something else to consider is the warranty. Some have a few years and some have lifetime (PNY & EVGA, I think).

And with both of those latter brands (eVGA and PNY), only selected cards have a lifetime warranty - namely, the eVGA cards that have their full model numbers ending in "-AR" and those PNY XLR8 line of cards. The other eVGA cards have only a two-year warranty while PNY's lower-level Verto line have a three-year warranty.

To the OP (Harry):

Excellent call on the 1GB GTX 460. However, the few results currently in the PPBM5 list on their Web site for the GTX 460 is slower than most other MPE-enabled cards because the currently available drivers for that card are not mature enough.

To everybody who responded to this thread:

IMHO a GTX 260 or higher is the best choice for CS5. Lower-class cards with 1GB or more RAM are not as good, but still provide a significant boost. The worst choice for CS5 are those NVIDIA cards which cannot use MPE's GPU acceleration mode (either due to the lack of CUDA or an insufficient amount of graphics RAM) because they perform even slower than ATi cards in software-only mode.

[Mike Greenberg]

Long time lurker, first time poster.

Currently have the Gigabyte 460 1GB model. Able to get it on sale for $199 recently. Not too loud at all.
Seems to be the best value per performance card that nvidia has produced in a while. Works great with MPE and the hack. Overall, probably the best card performance for price that I have owned in a long time.

[Harm Millaard]

Currently MPE is still in it's infancy and CUDA Toolkit 3.1 has just appeared.

What does that mean? First of all that at this moment in time there is no performance difference noticeable between a GTX-285 and a GTX-480, tested on the same system. This despite more than double the cores and double the width of the memory bus of the 480 over the 285.

At this moment the conclusion may be that the 460 is the most attractively priced MPE card with the same performance as the 480.

However, it is to be expected that when MPE get's more mature, it will use available cores more efficiently. Just like the threading on certain processes in CS5 that is far from optimal, causing much more latency in the communication between CPU - RAM - GPU - VRAM and back, especially on hexa cores.

Threading is a major factor for performance gains, both in the CPU and in the GPU. Therefore I expect that in time the 470 or 480 will be a much better choice than the 460.

[Jon Larson]

Harm,

I hope you are right about future optimizations improving the performance of these video cards, but do the 470/480 offer any other meaningful architectural advantages over the 460 beyond the "processor core" count? I know the memory bandwidth on the 480 is much higher Do the benchmarks show memory size/bandwidth (beyond the previously discussed limit) really matter for performance?

FWIW, here's a little info I gleaned off a simple search on Newegg:

GTX 260 896MB: 216 cores $180 $0.83 per core
GTX 285 1 GB: 240 cores $300 $1.25 per core
GTX 460 1 GB: 336 cores $230 $0.68 per core
GTX 470 1.2GB: 448 cores $350 $0.78 per core
GTX 480 1.5GB: 480 cores $450 $0.94 per core

[Randall Leong]

Originally Posted by Randall Leong

To everybody who responded to this thread:

IMHO a GTX 260 or higher is the best choice for CS5. Lower-class cards with 1GB or more RAM are not as good, but still provide a significant boost. The worst choice for CS5 are those NVIDIA cards which cannot use MPE's GPU acceleration mode (either due to the lack of CUDA or an insufficient amount of graphics RAM) because they perform even slower than ATi cards in software-only mode.

To the first two sentences in the above quote, it also depends on the CPU used. After all, it is way overkill to put in a GTX 480 in a Core 2 Duo system because the GPU would have been significantly faster than the CPU/system RAM subsystem. If CS5 were to run on such an imbalanced system, the GPU would then have to wait for the CPU to catch up.

On the other hand, some of the low-end GPUs have too slow of a graphics memory subsystem (read: too low of a memory bandwidth) to be of any good in MPE. Those cards which use 1GB of DDR2 memory showed no performance improvement (or even a performance drop) versus software-only mode. Hence, those bottom-of-the-line cards are not recommended for CS5 even if they have 2GB of DDR2 graphics RAM.

As for the last sentence, it all boils down to the different optimizations in the different manufacturers' drivers. In my testing, ATi's drivers are actually significantly faster than NVIDIA's drivers in MPEG-2 SD encoding performance but moderately slower in H.264 HD encoding performance.

[Harm Millaard]

Randall,

Just today I have updated both the Interpreting Results page and the Background Information page, to give some more information about the differences between hardware and software MPE when encoding. It may help you understand what causes the differences between ATI and nVidia cards and MPE on/off results.

However, keep in mind the maximum quality settings when using CUDA/MPE.

PPBM5 Benchmark

[Steve Kalle]

Harm,

Can you point me to where it says that Maximum Quality is engaged when using hardware acceleration because MRQ is not engaged by default or in PPBM5. Also, using MRQ greatly increases quality as well as rendering time, but it is worth every extra second of rendering time.

[Adam Gold]

So I went with the 480 in hopes there will be future enhancements down the road. I swapped the new 480 into my Supermicro Xeon system and re-ran the PPBM5 rendering test only. I did not run the encoding tests as I assumed these would be unaffected by the new card.

With MPE SW acceleration only the rendering time was 199 secs, virtually the same as with my older ATI card.

With GPU Acceleration turned on after the hack, the rendering time was 20 secs – a tenfold increase. Remarkable.

Good God, that 480 is a huge beast, with what looks like chrome exhaust pipes wrapping around the body and a fan that extends over the CPU heat sinks (or something – it barely fits into my huge Supermicro case). It takes THREE 6-pin power supply cables, as the only unused 8-pin connector from my PSU was the wrong type. Fortunately EVGA included an 8-pin to Two 6-pin adapter cable, and all fits and seems to be working well. At least there is no apparent smoke that I can see or smell.

By the way, a simpler form of the hack is to simply find the txt file as directed, and replace the “285” in “GeForce GTX 285” with “XXX”, where XXX is your model number. Reduces the chance of typos this way and eliminates the need to go into that other exe file and all that nonsense. I suppose you could do the same with the Quadro card numbers as well. We observed from other posts that even a capitalization error renders the hack ineffective.

[Randall Leong]

Originally Posted by Adam Gold

So I went with the 480 in hopes there will be future enhancements down the road. I swapped the new 480 into my Supermicro Xeon system and re-ran the PPBM5 rendering test only. I did not run the encoding tests as I assumed these would be unaffected by the new card.

With MPE SW acceleration only the rendering time was 199 secs, virtually the same as with my older ATI card.

With GPU Acceleration turned on after the hack, the rendering time was 20 secs – a tenfold increase. Remarkable.

In my experience with the GTX 470 (versus the HD 5770) the MPEG-2 encoding speed was actually slower with the 470 than with the 5770 but the H.264 encodes were faster. Also, with MPE SW acceleration only the 5770 was slightly slower in timeline rendering speed than the 470.

Also, I found differences between the 257.21 and the 258.96 drivers in my testing with the 470. The 257.21 drivers were a bit faster in encoding but the encodes suffered in quality, with black or green squares ruining the encodes through the 257.21s.

And of course the difference between MPE in software-only mode and MPE in GPU-accelerated mode with the 470 is roughly ten-fold.

But while the 470 is a great card for PPro CS5 use, it is not really needed there (though a system with a highly overclocked hexacore CPU would have benefitted from the 470 over the GTX 2## series GPUs). I could have gotten nearly the same level of performance by going as low as a 1GB GTS 250 (at least with my particular system's CPU). And had I continued to use my secondary system's Core 2 Quad Q9450 for CS5, I could have gotten away with even a 1GB DDR3 or GDDR5 version of the GT 240.

[Peter Chung]

I posted this in another thread and thought it would add to the discussion here:

Here's an excerpt from David Knarr's well-written article:

Originally Posted by http://www.studio1productions.com/Articles/PremiereCS5.htm

To give you an example of the performance difference between DDR3 and DDR5 memory, we ran some tests using a GT240 video card with 1 gig of DDR5 memory and a GT240 video card with 1 gig of DDR3. The video card with the DDR5 memory gave us around a 45% speed increase over a GT240 with DDR3 memory.

We also tested the GT240 with DDR5 memory against a GT250, GT260 and a GT 430, with DDR3 memory, The GT240 with DDR5 memory was faster in every test, even though some of the video cards had more CUDA cores. This was due to the GT240 having DDR5 memory and the fact that the Mercury Playback Engine isn't using all of the CUDA cores.

We are trying to find out exactly how many CUDA core the Mercury Playback Engine is using. I have heard rumors that it only use around 100 CUDA cores. Now remember, this is only a rumor. But from the tests we have run and what a few other people are reporting, I would have to say this may be true.

What I get from this is that the most important criteria in choosing a card for MPE playback right now (not taking into consideration potential in the future) are: 1) DDR5 memory, and 2) at least 100 cores (96 cores is close enough).

A logical conclusion, if these statements are true, is that having more cores only provides marginal improvements.

Elsewhere in his article, he says:

Right now, you will get the same performance across all the GTX 400 series of video cards. Right now, the performance is about the same as the GTX240 to GTX 285 video cards.

It would be great if you could chime in, David and Harm, with your experiences and if there are any considerations for buying a more expensive card than a GT240 with DDR5 memory right now or in the very near future.

It would be very exciting if a GT240 provides close to the same performance as a GTX480 when it comes to Mercury Playback. I don't play games and am just interested in Premiere CS5 performance. Also, I believe the GT2xx series is compatible with OSX, whereas there are no Fermi (4xx series) drivers (yet) for OSX.

Thanks in advance!

[Randall Leong]

Originally Posted by Peter Chung

I posted this in another thread and thought it would add to the discussion here:

Here's an excerpt from David Knarr's well-written article:


What I get from this is that the most important criteria in choosing a card for MPE playback right now (not taking into consideration potential in the future) are: 1) DDR5 memory, and 2) at least 100 cores (96 cores is close enough).

A logical conclusion, if these statements are true, is that having more cores only provides marginal improvements.

Elsewhere in his article, he says:


It would be great if you could chime in, David and Harm, with your experiences and if there are any considerations for buying a more expensive card than a GT240 with DDR5 memory right now or in the very near future.

It would be very exciting if a GT240 provides close to the same performance as a GTX480 when it comes to Mercury Playback. I don't play games and am just interested in Premiere CS5 performance. Also, I believe the GT2xx series is compatible with OSX, whereas there are no Fermi (4xx series) drivers (yet) for OSX.

Thanks in advance!

The 1GB DDR5 version of the GT 240 or the newer GTS 450 would at least suffice for most systems, especially the "consumer" systems that are to be used for CS5. The only way that the 1GB DDR5 version of the GT 240 might be limited would be if you're stuck with an early Intel Core 2 Duo or an AMD CPU with fewer than four cores.

[Steve Kalle]

But what about the ram's Bit? Both of the GT 240s only use 128-bit whereas other cards go up to 448-bit with DDR3 ram, and this increases bandwidth to similar, if not, greater than DDR5/128-bit.

[Randall Leong]

Originally Posted by Steve Kalle

But what about the ram's Bit? Both of the GT 240s only use 128-bit whereas other cards go up to 448-bit with DDR3 ram, and this increases bandwidth to similar, if not, greater than DDR5/128-bit.

This is because in my testing MPE barely takes full advantage of even 128-bit graphics memory, let alone 448-bit.