Any GTS 450 Tests Yet?

[Peter Chung]

Originally Posted by David Knarr

Hi Peter,

I am chimming in....

I am still running tests, but what I have found is, on the SAME system testing a GT240 with DDR5 memory gave me the same performance as a 460 and 480. (I was testing on an AMD Quad Core system at 2.9 Ghz. The main reason is the Mercury Playback Engine does not use all of the CUDA cores that are availble on the higher end cards.

In a test I was running last night ( and I haven't posted this yet on our website) when running on a AMD 6 core system, we did see a slight improvement between the 240 and a 460 and 470 card. It wasn't that much of an improvement, about 10 to15%. To me it wasn't worth the expense of the 470 card.

However, with all of that said, if Adobe in the future changes the Mercury Playback Engine to use more CUDA cores, then of course the 460 and 470 will be faster cards.

Also, as a side note. DO NOT bother with the new GT 430 card, it only has DDR3 memory. I tested this card last night also and the GT240 card with DDR5 memory is about 40% to 50% faster.

All test were run with the lastest drivers 258.96 WHQL

Thanks for chiming in, David. Again, thank you for writing such an extensive article on the MPE performance in CS5.

I am doing research for a new build, switching back from Mac to PC. My build will most likely be a Core i7 950.

A GT 240 1GB DDR5 can be had for as low as $76 and I've seen the GTS 450 1GB for $105. Those are pretty hot deals, especially if their performance is comparable to a $300 GTX 470 or $200 GTX 460.

My main preference for wanting to go for the GT 240 vs the GTS 450 is in case I miss my Mac and want to turn it into a Hackintosh. To my understanding, there are no Fermi drivers for Mac so the 4xx series wouldn't work in the Hackintosh, if I go that route. The GTS 450 is probably the best value, though, and some room for improvement if Adobe improves CUDA core usage.

How does the GT 240 performance compare with the GTS 450? Is the GTS 450 limited significantly by its 128bit memory bandwidth? It seems the GT 240 also has a 128bit memory bandwidth...

Note: I do check back and refer to your article often and it looks like it gets updated fairly often. It would be helpful to know what you've updated to your article so I can keep current more easily. Thanks again!

[Randall Leong]

Originally Posted by Peter Chung

Thanks for chiming in, David. Again, thank you for writing such an extensive article on the MPE performance in CS5.

I am doing research for a new build, switching back from Mac to PC. My build will most likely be a Core i7 950.

A GT 240 1GB DDR5 can be had for as low as $76 and I've seen the GTS 450 1GB for $105. Those are pretty hot deals, especially if their performance is comparable to a $300 GTX 470 or $200 GTX 460.

My main preference for wanting to go for the GT 240 vs the GTS 450 is in case I miss my Mac and want to turn it into a Hackintosh. To my understanding, there are no Fermi drivers for Mac so the 4xx series wouldn't work in the Hackintosh, if I go that route. The GTS 450 is probably the best value, though, and some room for improvement if Adobe improves CUDA core usage.

How does the GT 240 performance compare with the GTS 450? Is the GTS 450 limited significantly by its 128bit memory bandwidth? It seems the GT 240 also has a 128bit memory bandwidth...

Note: I do check back and refer to your article often and it looks like it gets updated fairly often. It would be helpful to know what you've updated to your article so I can keep current more easily. Thanks again!

I can explain why the GT 240 DDR5 did as well as it did: MPE (as currently implemented) barely takes full advantage of current 128-bit DDR5 graphics memory (as used on most mainstream graphics cards these days). The GT 240, like the GTS 450, has a 128-bit memory bus.

[David Knarr]

Brad, yes I check this morning and the latest driver is now 260.89. I just got down loading it on all of our computers.

Peter, right now the 450, 460, 470, 480 all perform pretty much the same with Premiere CS5 (when tested on the same computer, an AMD X4 and X6 systems.) Who knows about the future if Adobe starts using more CUDA cores.

I haven't got access to an Intel I7 system to run the tests on myself, but from the emails I have recieved, it doesn't sound like there is a whole lot of difference between the video cards on the I7 CPUs.

Peter, the 240 and 450 performed the same on my systems here. Both are 128bit bandwidth.

Randall, I agree with your statement "If you have an older or slower system, such as a Core 2 Quad or a quad-core AMD system, just pick the least-expensive NVIDIA card with at least 1GB of DDR5 memory that you can find (if you're going to run just CS5 on that system).

One person did email about having a I7-930 system that was overclocked. He didn't see any difference between the 460 and the 480 with Premiere CS5. These were the only two video cards he had to test. He is trying to get his hands on the GT240 or 250 to see what the performance difference is like.

Question to everyone -- Does anyone know how many CUDA cores Adobe is using with Premiere CS5? I have heard several rumors that it is only around 100 CUDA cores and that is why the performace isn't that different bewteen most of the video cards when tested in the same system.

[Mikel Arturo]

What's the successor to the Nvidia GT240?

All the series have been updated to 4xx, but I don't see the update of the 240.

More or less:
- 270 to 470
- 260 to 460
- 250 to 450
Etc. But, what about 240? No Fermi upgrade?

I ask this because is a one year old card (the last of the 2xx series) and all the versions have been upgraded to Fermi chops. Is coming a 440 or similar card?

[Peter Chung]

There is the GT430 but I wouldn't recommend it because it uses DDR3 memory instead of DDR5. Both the GT430 and GT240 have 96 cores but you can get the GT240 with DDR5 memory.

[Harm Millaard]

Randall,

If you have a look at Latest News you will see I mentioned you explicitly. If you want me to remove that sentence, please say so, but I think it shows how you can use a system that really delivers performance without spending all future inheritances now. It may not be the top performer, but per $ spent, it now ranks in second place after my own system, albeit at the very left of the horizontal axis around 30%, which is not visible on the scale used for the top 25 performers.

[Randall Leong]

Originally Posted by Harm Millaard

Randall,

If you have a look at Latest News you will see I mentioned you explicitly. If you want me to remove that sentence, please say so, but I think it shows how you can use a system that really delivers performance without spending all future inheritances now. It may not be the top performer, but per $ spent, it now ranks in second place after my own system, albeit at the very left of the horizontal axis around 30%, which is not visible on the scale used for the top 25 performers.

Thanks for the mention. It does mean that not everyone needs a top-performing system for his or her own needs.

By extension, it also means that a Socket 1156 system (i7-8xx) can be used to achieve similar results as long as one does not overtax the system with an excessive number of PCI-e expansion cards (as the LGA 1156 CPU itself has only 16 PCI-e 2.0 lanes running at full bandwidth and the P55 chipset adds an additional eight PCI-e 2.0 lanes which are restricted to PCI-e 1.0 bandwidth). The caveat: Any hardware RAID controller added to a P55 system will either be forced to run at PCI-e 1.0 x4 bandwidth or drop the graphics card's slot to PCI-e 2.0 x8 mode. That will limit the future expansion capability of that platform since only three PCI-e 1.0-bandwidth lanes will then be available (after accounting for the onboard LAN which on most motherboards eat up one PCI-e 1.0-bandwidth lane). And forget about P55 or 1156 if you also want to use USB 3.0 and/or SATA 6 Gbps: Current onboard solutions eat up up to four additional PCI-e lanes. On Gigabyte's P55A-UD3 motherboard, the use of a hardware RAID card will force the onboard SATA 6 Gbps and USB 3.0 controllers to use the Turbo mode, which will drop the primary PCI-e 2.0 x16 slot to x8 mode.

[Randall Leong]

Originally Posted by Harm Millaard

Randall,

If you have a look at Latest News you will see I mentioned you explicitly. If you want me to remove that sentence, please say so, but I think it shows how you can use a system that really delivers performance without spending all future inheritances now. It may not be the top performer, but per $ spent, it now ranks in second place after my own system, albeit at the very left of the horizontal axis around 30%, which is not visible on the scale used for the top 25 performers.

Again, Harm, thanks for the acknowledgement.

I have since rebuilt that system with a RAID 0 array and a better CPU cooler (and a new name). I retested that system, and submitted its PPBM5 result. Its 317-second total time shows you that a properly tuned i7-9xx system with only 6GB of RAM can outperform a poorly tuned i7 system with 12GB or 24GB of RAM.

[Peter Chung]

Randall's results also show that a GT240 (I'm assuming with DDR5) can hold up against a GTX470(!)

What else did you tune besides higher overclock, RAID0 array, new CPU cooler (which are you using?), and 5.03 upgrade?

Also, what did you change from Steamer 3.7 to Steamer 3.67? Did the upgrade from 5.01 to 5.02 cut the results a whole minute or was there something else?

And on a side note, have you tried running a hackintosh on your Steamer? The UD3R seems to be very popular for that and I'm thinking of using it for my next build.

Thanks!

[Randall Leong]

Originally Posted by Peter Chung

Randall's results also show that a GT240 (I'm assuming with DDR5) can hold up against a GTX470(!)

What else did you tune besides higher overclock, RAID0 array, new CPU cooler (which are you using?), and 5.03 upgrade?

Also, what did you change from Steamer 3.7 to Steamer 3.67? Did the upgrade from 5.01 to 5.02 cut the results a whole minute or was there something else?

And on a side note, have you tried running a hackintosh on your Steamer? The UD3R seems to be very popular for that and I'm thinking of using it for my next build.

Thanks!

The upgrade from 5.0.1 to 5.0.2 improved my overall results by a minute. Also, I changed from 3.7 to 3.67 because the CPU was running at 3.738 GHz when I ran 5.0.1 compared to 3.675 GHz when I ran 5.0.2.

On the system with the 6GB of RAM, I did no other tuning besides what I stated above. 5.0.2 and 5.0.3 performed virtually equally to one another. I switched from the stock Intel boxed CPU cooler (which limited my overclock to 3.36 GHz because any higher resulted in an internal CPU core temperature that exceeded 85°C in Prime95 and IBT) to a Cooler Master Hyper 212+ (which kept the CPU temperature to the upper 60s C despite the higher overclock to 3.677 GHz, and my i7-920 could not go higher than that without requiring a Vcore increase to higher than 1.225V, which is the upper end of Intel's nominal operating voltage range).

[Randall Leong]

Since then, I have noticed three very slow results in MPE from the GT 240 in two different systems - 63 seconds on one system and 107 seconds (CPU at stock speed)/92 seconds (CPU overclocked) on another system. I am suspecting that either they have driver issues or are using GT 240s with only DDR3 memory. This compares to the 10 to 12 seconds that the DDR5 version of that same card can achieve in a properly tuned system.

[Randall Leong]

Originally Posted by Randall Leong

Since then, I have noticed three very slow results in MPE from the GT 240 in two different systems - 63 seconds on one system and 107 seconds (CPU at stock speed)/92 seconds (CPU overclocked) on another system. I am suspecting that either they have driver issues or are using GT 240s with only DDR3 memory. This compares to the 10 to 12 seconds that the DDR5 version of that same card can achieve in a properly tuned system.

The system that had the 107/92 seconds was retested because some of the video clips were off-line when he originally tested the system. His new result was 13 seconds in MPE mode.

[Rob Johnson]

Originally Posted by Steve Kalle

I just noticed that the GTS 450 with 1GB vram and 192 cores is available for only $130. I am ordering 2 more HP workstations (Z400 & Z600) in a week and I was considering the 460 until I saw the 450, but I would like to know if there is a noticeable performance difference within Premiere & AME.

Hi Steve, I'm currently in the same situation as you were, and have also been debating between the GST 450 and GTX 460. But in light of the more-than-extensive information David Knarr has gathered and posted about the GT 240, I think I may just go with that one (1G DDR5 of course). From everything I have read, it seems to be almost right on par with a GTX 470 or 480 (!) where the CS5 MPE is concerned. I play no games on my vid editing box so that's no concern, and it will save me a PSU change-out. Just something to consider.

I supposed the only caveat would be the distinct, and most likely, possibility that PPCSx will use more CUDA cores in the future (it will be a great hook for an upgrade). As of now (aside from having a Xeon X5680, i7 980x system or greater) that is the only reason I can see for getting anything more than the GT 240.

[Randall Leong]

Originally Posted by Rob Johnson

Hi Steve, I'm currently in the same situation as you were, and have also been debating between the GST 450 and GTX 460. But in light of the more-than-extensive information David Knarr has gathered and posted about the GT 240, I think I may just go with that one (1G DDR5 of course). From everything I have read, it seems to be almost right on par with a GTX 470 or 480 (!) where the CS5 MPE is concerned. I play no games on my vid editing box so that's no concern, and it will save me a PSU change-out. Just something to consider.

I supposed the only caveat would be the distinct, and most likely, possibility that PPCSx will use more CUDA cores in the future (it will be a great hook for an upgrade). As of now (aside from having a Xeon X5680, i7 980x system or greater) that is the only reason I can see for getting anything more than the GT 240.

It's been over nine months since the last post in this thread, but I am going to update this (based on my own personal results) now that Premiere Pro CS5.5 is out and is up to 5.5.2 (as of this posting).

Simply put, the GTS 450 is where the performance in CS5.5 will start slowing down significantly, especially in HD-to-MPEG-2 SD transcodes: It has only 128-bit access to its DDR5 RAM. And the GTX 550 Ti, which effectively replaces the GTS 450, is just slower than a GTX 260 due to the former's 192-bit DDR5 memory bus (which puts its memory throughput just below that of the GTX 260 with its 448-bit DDR3 memory bus). As such, the GTX 550 Ti is acceptable if your system has a low-end i5 or below for a CPU - but I'd strongly recommend a GTX 560 or higher if you can afford it even if your CPU isn't quite up to speed.

As for the GTX 470, my particular GTX 470 only performs on a par with a top-end GPU from the generation that's previous to the Fermi series, the GTX 285. That's mainly because the Nvidia reference cooler for the GTX 470 could not prevent the GPU from throttling itself down in clock speed during MPEG-2 DVD encodes -- in other words, the GTX 470's reference cooler is ineffective.

Also, keep in mind that the faster or more highly overclocked your system's CPU is, the higher-end (faster) GPU you'll need in order to keep the system's performance in good balance. That's because the article referenced for the CUDA hack used an AMD Phenom II x4 CPU, which lacks support for the SSE 4.x instructions that Intel CPUs have and which Adobe Premiere Pro makes extensive use of. What's more, the Phenom II x4 only performs on a par with a higher-end old-generation Intel Core 2 Quad or a low-end dual-core Sandy Bridge i3 CPU overall in the PPBM5 tests under Premiere Pro CS5.5. A true quad-core i5 of the Sandy Bridge generation (such as the i5-2400) will easily outperform the AMD CPUs in this situation.