Premiere Pro CS6: GTX 660 ti ($300) vs. GTS 450 ($100) and other thoughts on upgrading HW

Question

I've been wanting to make some major upgrades to my hardware, but it just doesn't seem worth it yet...even after almost 4 years. I ultimately decided to "rent" a new video card and run some tests. Here is some background info on my upgrade though process and some results comparing the video card performance.
Disclaimer: I'm not a hardware expert, but I'm not completely clueless (I think). Your input/insight is welcome.
My system (purchasd 2/2009)
i7 920
GTS 450 (1GB RAM)
12 GB 1333 RAM
Samsung SATA II 128 GB SSD (OS/apps)
5x 1TB 7200 RPM drives in RAID 0 (with accompany slow/cheaper 2TB backup drives)
Some upgrade options I am considering
Sandy Bridge 3930 - but it's $560 w/o cooling and would require a new, more expensive motherboard, new ram, cooling, etc.
Ivy Bridge 3770, but I keep reading that that an overclocked 920 isn't that much different in perf (in fairness mine isn't oc'd). I did find a MB that would work for only $90. So I could make this upgrade for just under $400 (RAM would stay the same).
Wait for Haswell, but i could be another 9 months and it's supposed to only give maybe a 10% perf gain over IB. It's more focused on mobile - less power, integrated graphics, etc.
High-end Xeons are totally off the table. $/buck is waaaay too low.
Video card and benchmark reviews/problems
So I thought I'd first try getting a new video card. I see conflicting benchmarks. This site (the one that provides the CUDA.exe hack) notices very little difference between most GTX cards in perf for their benchmarks. The PPMB5 site shows significant differences between say the GTX 680 and lower end cards. But are these really accurate?
The GTX 680 is almost $500, so I opted for the 660 ti at $300 to see if I could get a noticable perf gain. It seemed like the best $/buck card and wouldn't require me to get a new power supply.
Another reason I wanted to do my own tests: None of the benchmarks I've seen actually mention the type of footage used. I care about footage from the Canon MKII-III, or similar footage. I definitely do not care about things like exporting to MPEG 2.
I did some very unscientific benchmarks, but they were real world for me. First my "problem" areas.
Performance problem areas
#1 - Time-lapses consisting of 1080 (height) JPEGs and 2160 (height) JPEGs don't always play smoothly (larger 2160s almost never do). I read adding more VRAM might help. The 660ti has 2x the RAM as my current video card.
#2 - Split screen sequences (up to 9 clips simultaneously) don't play smoothly.
#3 - Scenes where I speed up a clip to 1000x don't always play smoothly. (Although upgrading from CS5 to 6 actually seems to have solved this issue, I couldn't get it to repro any longer).
#4 - Export to h.264 could be faster. I do this a lot, but mostly because it's how I sometimes make proxies because of problems around #1-2 (works fine - used to use CineForm but it always crashed Premiere and these work for my needs). This is typically my final export as well for posting on sites like Vimeo.
#5 - Timeline rendering could be faster, although I don't do this a lot and if I do it's simple, not a bunch of crazy effects. E.g. use unsharp mask. This is pretty low pri for me though because I think timeline rendering is a bad idea. Once you do it, if you even move the clip you have to render again.
Some simple bottleneck analysis first:
Disk queue length sometimes is just over 1 on 1 disk in my RAID array during TL playback. Might slow things down slightly. Not an issue during export.
Processor never seems to get pegged in any case.
RAM is never maxed out, but it starts to go to Premiere limits (10 GB that I've set) after playing through several time-lapses (I'm just now noticing this). Choppiness starts well before RAM is even near that on some clips.
Tests/results:
NOTE: I do run the 660ti in a PCIe 2 x16 slot. Let me know if you think it would even matter to run in a PCIe 3.0 slot. My MB doesn't have one.
#1 Time-lapse smoothness - didn't improve with the 660. Moving the 1080 size JPEG TLs to my SSD did help some problem TLs play smoothly however.
#2 Split screen. Did a test with a 9-clip-at-the-same-time sequence. No improvement with the 660ti.
#3 Clips speed up 1000x - could not repro the problem now that I run CS 6 vs. 5 on either card.
#4 - Export to H.264 1080p @23.9x fps.
Export 5:30 clip of 5D MKIII footage + H.264 proxies:
GTS 450 - 9:14
660 Ti - 8:30
Export 1.5 minute clip of large time-lapses (JPEGs that are 2160 high):
GTS 450 - 9:35
660 Ti - 7:00
Export a 2 minute clip of just MKIII footage
GTS 450 - 2:45
660 Ti - 2:45
#5 Timeline render with simple image correction effect
Timeline render short 5D MKIII clip with unsharp mask applied:
GTS 450 - 1:10
660 Ti - 1:19
Conclusion:
The 660ti ($300) showed marginal improvements in exporting h.264 against my GTS 450 ($100) and did not address my other issues. Definitely not worth it for the type of work I do.
Moving my time-lapse JPEGs to an SSD helps play the 1080p versions back smoothly. The 2160p larger versions still lag. Maybe more RAM would help? They still start off choppy and then acquire more and more RAM, so not sure here. Maybe faster 1600 RAM? I don't know, I doubt it. I may have to just use 1080 versions or make proxies.
I don't see a pegged CPU much if at all, so upgrading to an Ivy Bridge 3770 doesn't seem like it'll help much if at all.
I did end up buying 2x256 GB SATA III SSDs (only $169 each) that I'll run current projects off of, or at least time-lapse sequences (RAID 0). My motherboard doesn't have an SATA III slots, however, so I won't see the full power of these, but not sure I'll need it. Again I'm not seeing a clear disk issue either from the perf monitoring.
I suspect many of these problems are still with the software and how it takes advantage of my hardware, but I'd love more insight.
Generally I make things work and I don't have any really painful bottlenecks, but I'm always up for perf improvements/doing things faster. It does look like I won't see any major breakthroughs, however by spending $400-$1000 bucks on HW upgrades.
Thoughts?
Luke
Blog  |  Photography  |  Vimeo

Answer 1

Thanks for the resonpose Harm. Inline.
Harm Millaard wrote: SYSTEM: It is an older system, about the same I had in the form of 'Harm's Beast', although I have a much beefier disk setup, more memory and OC'ed to 3.7 GHz, in combination with a GTX 480. Not much you can do about this system, apart from upgrading memory to 24 GB but the major drawback is that those investments will not carry over to a new system, at least not easily. [Luke] From your description of your system it sounds like 4 things could indeed be upgraded and carried over to a new system. 1) OC the processor (e.g. purchase a generic water cooler for ~$100), 2) Improve the Disk setup, 3) Upgrade the video card, 4) Add more/faster RAM.
I've seen in some benchmarks that an OC'd 920 is not so dissimilar to an OC'd 3770K. The latter is faster, but it isn't a huge difference. The larger question still remains - will any/all of these upgrades yield large performance gains and solve all/a higher percentage of my problems? Or do I have a decent sweet spot of a system and should wait for the software (e.g. MPE evolution in CS7-8) to catch up and take better advantage of what I have? Like I said from doing some rudimentary performance monitoring, I'm not seeing a pegged CPU (just a brief spike here/there), I'm not seeing disk transfer at capacity (although 1 disk has a slightly > 1 queue length at times), I'm not seeing in all cases over-utilization of memory, etc. (except higher RAM usage is seen albeit staggered for large JPEG time-lapse sequences, but I see choppiness well before RAM usage gets to 10 GB).
 VIDEO: You correctly point out that the GTX 680 shows in the MPE graph on the PPBM5 website much better results than other cards. But keep in mind that most 680's are used in new systems, often with the latest CPU's and fast memory. I'm convinced that a 680 is not noticeably faster than a 580, because they have the same memory bandwidth, but it looks that way because they are often accompanied by hexa core i7-39xx CPU's with large amounts of memory.  [Luke] Good point - potentially further evidence that the video card doesn't make a big difference? At least not enought to justify 5x the cost (e.g. $500 680 vs. $100 450).  This would be consistnet with what Studio 1 Productions has seen. The GTS 450 has a memory bandwidth of 86 GB/s, the 660 Ti has 144.2 GB/s, so the latter is significantly faster as you have shown in some of your tests. [Luke] The only test I would characterize closer to having a significant increase using the 660 would be exporting large JPEG time-lapses to H.264, where it was a good 27% faster. The rest seemed more marginal or did not change. TESTING: You don't mention to what format you exported and with what resolution and framerate. Hardware MPE will come into play when you have rescaling, frame blending, blurring and stuff like that occurring. If you export to the same frame size and frame rate as your source and no blurring occurs, then exporting is purely a CPU matter and the video card has no impact at all. [Luke] See above - H.264, 1080p @23.9x fps. General remarks: I personally consider your 5 disk raid0 setup as pretty risky. You have multiplied the risk of losing all data by a factor 5!. You have no redundancy at all. Even though it is fast, I expect your sustained transfer rates are less than 450 MB/s and when using a 9 clip split screen, it may be too slow with the limited memory and the old CPU you have. You have effectively one single volume for video related editing (apart from the OS disk) and while that makes for easy administration, it still entails the drawbacks of the half-duplex connection of SATA. It might be better to add a couple of HDD's in raid0 for media cache, previews and exports to avoid that limitation. You can always carry those to a new system. [Luke] Yes there is a higher level of risk, but with backups every 30 minutes during project work I yield cheap/easy perf gains for the cost of--at most--30 minutes of work. I've lost no work in the last 4 years, 1 drive failed once while on vacation and I replaced it easily. Anyway backup/data integrity is a different issue separate from performance which I'd like to focus on in this context.
I get ~420MB/s read with this array (mostly older, blue WD drives and Deskstars). I'm running out of space, so I just ordered 3 x 2 TB WD black drives to replace this with, expecting probably a similar transfer rate. Again though I'm not necessarily seeing disk being a bottleneck in perf mon aside from one disk who's queue length sometimes goes over 1, so we'll see if the newer black drives help.
I have ordered 2x256 SATA III SSDs to put min my time-lapses on as having them on my current primary SSD seemed to help in some cases.
 Sorry to be so harsh. [Luke] No worries, harsh is OK but I'm still not seeing a clear solution to some of my issues and I'm still not convinced a new system - short of a top-of-the-line SB or Xeon system (both of which are very $$) will be worth the upgrade. When 5 years had passed between my current system and my 2004 system, I feel like the upgrades were much more significant especially for bang/$.
Luke Humphrey
Blog | Photography | Cinematography