I just replaced the i3-2100 CPU with the i5-2400 CPU, and the total PPBM5 benchmark time improved by another 146 seconds - from 524 seconds (with the i3-2100) to 378 seconds (with the i5-2400). This time, for some strange reason the disk I/O time improved from 155 seconds with the i3 to only 105 seconds with the i5. But the biggest improvement came in the H.264 Blu-ray time - from 147 seconds with the i3-2100 to only 98 seconds with the i5-2400. (And yes, there is another tangible improvement in both the MPEG-2 DVD and the timeline export/render tests, with MPEG-2 going from 213 seconds to 168 seconds and the MPE render time going from 9 seconds to 7 seconds.)
As such, if I were building a new low-budget build that will see occasional use of Premiere Pro CS5.5, I am now recommending an i5-2400 with 8GB of RAM, a GeForce GTX 560 card and at least two disks as the absolute minimum system configuration to run PPro CS5.5 to most users' satisfaction. This configuration would have performed roughly equal to that of an otherwise identical system but with a non-overclocked i7-920 CPU, an X58 motherboard and 12GB of RAM.
By the way, had I kept the GT 240 (which is now on the verge of failure because it started BSODing my system occasionally just starting Windows), the total PPBM5 benchmark time with the i5-2400 would have been almost as slow as on the i3-2100 primarily due to the MPEG-2 DVD times being almost equal in both systems due to the limited GPU performance.
Now that the auxiliary editing system is set, all it needs now are a new motherboard (the DH61CR used in the system is much less than ideal since it lacks aid0 support and has only four SATA ports with the two primary ports being completely blocked by a graphics card with a double-slot cooler, and that those two ports are much too close to one another to allow the use of two right-angle SATA plugs) plus a Blu-ray burner.
Realistically, the GTX 550 Ti would be the best match for a Sandy Bridge i3. The GTX 560+ GPUs really need a true quad-core or hexa-core i-series Intel CPU to perform at their fastest.
As for BFTB (Bang For The Buck), I would definitely say that the i5-2400 build with the GTX 560 card and an upgraded PSU is a significantly better BFTB (in terms of total system cost, not just the cost of the core components) than any i3 build with any GPU or PSU even though today's relatively high prices for hard drives make this particular i5-2400 build $875 (versus the i3-2100 system's $775 price in its most appropriately performance-balanced configuration) - with one 320GB to 500GB OS drive and one 1TB secondary drive.
After reading your two posts, I ran some tests today on my system which is:
An HP computer running an AMD Phenom II X4 (quad core) computer running at 3.2 Ghz with 8 gigs of RAM and dual Samsung 7200 rpm SATA 3.0Gb/s hard drive. The computer was running Window 7 Home edition. All of the video cards had DDR5 memory.
I used the PPBM5 for CS5, since I had that on the computer already. I was using Premiere CS5.5 with all the latest updates.
When I did a direct export, by selecting File, then Export, then Media. I used the following settings during the Export. Format = MPEG2-DVD, Preset = NTSC High Quality Widescreen, I then unchecked Export Audio and I unchecked Use Max Render Qualtiy, Use Previews and Use Frame Blending.
Here are my results:
In Software Mode = 387 seconds
In GPU Mode :
GT240 = 281 seconds
GT440 = 275 seconds
GTX470 = 230 seconds
GTX545 = 258 seconds
GTX550 Ti = 246 seconds
GTX570 = 193 seconds.
When I used the AME, by selecting File, then Export, then Media and then Queue. I set everything to the same as I did above. When I exported in GPU mode on the GT240 it took 823 seconds!!!
I know you and Bill are reporting that on your systems, with a low end card, it was faster to export to MPEG2-DVD file when you used the Software mode.
But I am not seeing that on my system or the way I am exporting the file.
Got any clues why???
The low-end card Bill tested was a GeForce 9500 GT: That card has only 32 CUDA cores and 1GB of slow DDR3 RAM. Compare that to your 96-CUDA-core GT 240, which is equipped with 1GB of faster DDR5 RAM. (Bill's 9500 GT took a whopping 783 seconds on even an Intel i7-2600K that's been overclocked to 4.4GHz, while the GTX 560 Ti 448-Core took only about 68 seconds. On that same PC, the GTX 550 Ti took 145 seconds.)
In addition, the AMD Phenom II x4 CPU is slow by modern standards: In CS5.5, it performs only on a par with a higher-level Socket LGA 775 Intel Core 2 Quad as well as a newer dual-core Sandy Bridge i3-21xx series CPU but significantly slower than a relatively low-end Intel quad-core Sandy Bridge i5-2400 CPU. The AMD CPUs prior to the FX series also lack full support for the SSE 4.x instructions that Intel's recent CPUs have (and which the Adobe Creative Suite programs make fairly extensive use of). No wonder why your results with the higher-end cards have been limited by the CPU.
As for my own test, the MPEG-2 DVD times were via AME. (The PPBM5 benchmark instructions specifically call for using the AME, not exporting directly from Premiere, for the encoding tests.) On my previous i3-2100, it took 311 seconds with the GT 240 versus 213 seconds for the GTX 560. Your 823-second result via AME with the GT 240 just shows you that HP pre-installs so much bloatware that actually makes even an i7 perform slower than even a Core 2 Duo on their home PCs. (This can be checked by opening the Task Manager and then clicking the "Processes" tab: Most pre-built PCs from the major brands such as HP have hundreds, if not thousands, of processes going on, compared to my auxiliary PC build that currently has only 55 processes going on.)
By the way, Bill's software-only MPEG-2 results were incorrect: Either the project still used MPE GPU acceleration-enabled mode or he did not bother to re-save the project in MPE software-only mode and re-run the MPEG-2 DVD portion of the test.
I retested my auxiliary editing system now equipped with an i5-2400 CPU but with the GT 240 DDR5. The MPEG-2 DVD encoding performance is even slower with MPE GPU acceleration than without: 278 seconds with GPU acceleration versus only 227 seconds in MPE software-only mode. However, the H.264 Blu-ray encoding performance actually improved with GPU acceleration than without - only 113 seconds with GPU acceleration versus 183 seconds in MPE software-only mode. This means that (at least with the i5-2400) the GT 240 is a bit too weak to handle both scaling/downsizing and MRQ (which is always enabled in MPE CPU accelerated mode) simultaneously.
I also performed a control run of PPBM5 with only the i5-2400's integrated Intel HD Graphics 2000 and no discrete GPU, and found that it actually performed a few seconds faster overall than that same PC with the GT 240 with MPE GPU acceleration disabled.
My results confirm that the AMD CPU in the HP system that you have been using is simply too weak to take much if any advantage of the higher-end GPUs. In fact, with such a slow CPU, it's not surprising that any plain old GeForce GPU with sufficient VRAM (at least one that's equipped with DDR5 VRAM) is faster than no Nvidia GPU at all (or one with too little VRAM, which cannot use MPE GPU acceleration at all) on that particular PC.
Message was edited by: RjL190365
To clarify things, first all of my systems has a New install of Windows 7 Home. from Microsoft. No bloatware. I always buy a fresh copy of the OS for each computer I run. I can't stand all the bloatware HP puts on their computer.
My HP AMD system that I am using for testing only 58 processes running.
I know the AMD in not the best system for Premiere due to the lack of SSE4.1, but my main use of all of my computer's is for animation work. I have a buddy that has an Intel based system with the same specs as mine and for rendering animation in Lightwave the systems perform about the same on the same animation file. So it is less expensive to use AMD on the rendering systems.
I use Premiere for final testing of the animations, for correct looping and to make small previews for my website. The AMD computer I have work just fine for what I use them for.
Today I ran a few more tests using AME in with both GPU and Software mode.
Here are the resulsts using the GT240:
AME GPU 823 seconds
AME Software 911 seconds
First Question - On my test system I am showing the the GPU does help when running through the AME. Any clue why I would be seeing that, when you and Bill are reporting the opposite?
Second Question - When I use direct export, why is that coming in SO much faster than using AME? I have double checked and all of the setting are the same. I remember something about this being discuss once before but I can't find it.
Direct Export GPU 281 seconds
Direct Export Software 387 seconds
You stated "with such a slow CPU, it's not surprising that any plain old GeForce GPU with sufficient VRAM (at least one that's equipped with DDR5 VRAM) is faster than no Nvidia GPU at all". However, I am not seeing that. I am seeing a performance improvement with the GPU over Software mode as I stated in my earlier post.
After todays test with the AME, I am still seeing an improvement over the Software mode.
I just now got off the phone with another guy who has an Intel Core2 Quad Processor Q9550 (he believes it's the Q9550) and he ran the benchmark test using the same version I am, on a system with no bloatware (61 processes) and with a GT440 card.
He said, using the Direct Export and in GPU mode, was 102 seconds faster than the Software mode.
When he use AME, the GPU mode was faster than the software mode by 145 seconds.
Sorry he didn't have the times, only the time difference between the GPU modes and Software modes.
He also said that Direct Export was about 4 times faster than using the AME.
Third Question - Any reason why is is also seeing direct export is so much faster than AME?
Thanks for taking the time to answer my questions.
Gotcha, Dave. The Q9550 is, if anything, just about as slow overall as the average AMD quad-core. And in my own testing I have already established (with a GT 240 running CS5 version 5.0.3 in GPU mode) that a system with a Core 2 Quad Q9450 quad-core CPU is already nearly 100 seconds slower than an otherwise identical system with a Sandy Bridge i3-2100 dual-core CPU.
Still, that does not explain how you got 823 seconds with the GT 240 on that AMD quad-core while I got only 311 seconds with the GT 240 on my i3 dual-core (AME queuing was used in both instances)...
And I did run my i3-2100 with the GT 240 in software-only mode: It needed nearly 370 seconds in the MPEG-2 DVD test. This tells me that although the GT 240 is a good enough choice for the lower-end CPUs such as the AMDs and the i3s or older-generation CPUs such as the Core 2 Quads, it would definitely hold back the performance of a quad-core i5-2400 or higher (and especially an overclocked i7-2600K or an i7-39xx CPU) - especially since my own testing on my current i5-2400 auxiliary editing rig has proven that with the GT 240 it actually performed about 50 seconds slower in GPU mode than in software-only mode when used in combination with that particular CPU (and that this difference in favor of software-only mode would have only widened with a faster CPU than this i5).
So, enjoy your 96-core GPU. Just be aware that the desktop Sandy Bridge i5-2300 or 2400 level is the CPU point where the CPU starts to overtake the 96-core GPU in absolute performance. (Although of course, MRQ is always enabled in GPU mode whereas it's turned off by default in software-only mode.)
By the way, have you heard of the question "When will pigs fly?"?
You said "This tells me that although the GT 240 is a good enough choice for the lower-end CPUs such as the AMDs and the i3s or older-generation CPUs such as the Core 2 Quads"
That's what I was thinking. Thanks for testing the GT240 in your system. I didn't want to just assume that's what was going on.
BTW, I have all of the video cards I listed above and I do enjoy my 96-core GPU. It does the job for what I need it for.
I will be pickup up a Intel I7-2600 system shortly, so I will be dropping in my GTX570 into it.
Do you have any clue as to why using the AME it is so much slower than direct export on the MPEG2-DVD??
If I have time tomorrow, I will do an un-install of CS5.5 and run the CS5 cleaner on the system, and reintall CS5.5 to see if there any improvement with the AME.
One more thing, do you mind if I put some of your info in your last post in my article on my website?
I was using a Q9550 system with an HD4850 with 2Gb before switiching to my X79 3930K system, and it actually worked reasonably well. I had to render the Timeline to get a decent preview, but it didn't take so very long, and applying edits and just about everything other than previews and exports was fairly comfortable. If I was earning my living with PP though, it would be a different story.
I don't mind.
By the way, I replaced the H61 motherboard with a new Gigabyte Z68 mATX motherboard (GA-Z68MA-D2H-B3, to be specific). That's because I found the H61 motherboard too limiting for an editing rig: It has no RAID support at all, nor does it natively support more than four SATA devices (three disks plus one optical drive). Worse, most H61 motherboards have no USB 3.0 or eSATA ports at all. And the Intel DH61CR board that I used has one additional limitation: The double-slot cooler on higher-end Nvidia GeForce GPUs (GTX 550 Ti and above) completely block the first two primary SATA ports, forcing one to use two "left-angle" SATA plugs, which in reality are right-angle plugs designed upside-down (since the two SATA ports are so close to one another) just to accommodate both SATA devices on the primary ports.
Like you with your Q9550 system, I don't have any problems on the quad core AMD system. It's fairly fast and easy to edit on. Depending on how many tracks I use, I may or may not have to render the timeline to get a smooth playback.
But, I totally understand Randall's position on the AMD and lower end systems. I just don't need it for my application.
Just wanted to let you know, I wiped my hard drive clean, reloaded Windows 7 and CS5.5 and nothing else. Now when I use AME GPU via Queue, it now renders at 459 seconds instead of 823 seconds like before.
However, direct export with GPU, is still faster. It is coming in at 280 seconds. So something is going on with AME.
It took me nearly three months to respond to this, but I think there is a serious bottleneck somewhere in the systems that you are using. Especially since your systems are using just a single hard drive for absilutely everything - OS, programs, media cache, pageing file, media, projects, renders and output - all on the exact same single drive. This will kill performance in queued (AME) mode.
I retested my auxiliary i5-2400 system with multiple drives using the direct export. I cannot reproduce your findings: In my case, the direct export is actually slower in both the MPEG-2 DVD and the H.264 Blu-ray tests than the AME-queued mode. For example, the MPEG-2 DVD test in direct export mode took all of 209 seconds versus only 153 seconds in queued (AME) mode. And H.264 direct export took a few seconds longer than AME did in my system, as well.
I did not search this whole thread but if you are using PPBM5 for this testing there is an anomaly in that our test the MPEG sequence is repeated three time to make it more measurable, This worked fine in CS5 but not good in CS 5.5. Unfortunately we did not find this out until after CS5.5 came out and so not to invalidate hundreds of tests we decided to just live with it. I have taken several real video projects shot with AVCHD 1920 x 1080 and my normal output is MPEG2-DVD 720 x 480 and encode both direct export from Premiere and with the sequences selected like we do in PPBM5 only in AME. I have yet to find a difference in encoding times between the two methods.