Here Build a Desktop Video Editing PC
-3 price level ideas in http://www.pacifier.com/~jtsmith/ADOBE.HTM
-what PC to build http://forums.adobe.com/thread/947698
-2 how to build videos http://forums.adobe.com/thread/1104182
-another video http://forums.adobe.com/thread/1145366
-another video http://forums.adobe.com/thread/1132363
-planning & Building http://ppbm7.com/index.php/intro-part-1
Thank you John. I have looked at may resources like that, I don't need help building a PC or choosing the hardware. What I want to know is what I'm going to get more out of Adobe's software with: CPU or a GPU power. I'll take a look at those resources to see if there's anything of value in there, appreciate your post.
Edit: I see that first resource goes into a ton of depth. I'm going to have a lot of reading to do, but as an amateur I would still love some general commentary on where I can get the most bang for my buck for now. I'm 2 months out from building my machine so this is the preliminary phase.
GPU acceleration requires both. The CPU has to process the data ie decode it and manage ram buffers for the GPU acceleration to work. This means the CPU starts the process pipeline. Right now the GPU's are not the bottleneck. The CPU's are so you need atleast a Quad Core CPU with Hyperthreading and a 750GTX card together for most standard base workflows. GPU acceleration uses allot of ram so 16GB is really minimum with most HD workflows otherwise performance may start to decline.
I have started to created, for myself, a build that would be in you need and price range.
Keep in mind that this is geared toward a OC system, if you're not really comfortable with this you may change the MB and reduce the cooler.
G.SKILL Ripjaws X Series 16GB (2 x 8GB) 240-Pin DDR3 SDRAM DDR3 2133 (PC3 17000) Desktop Memory Model F3-2133C9D-16GXH Keep in mind that Windows Home will be limited at 16 gigs (at least W7 Home is)
MSI Z97 MPOWER LGA 1150 Intel Z97 HDMI SATA 6Gb/s USB 3.0 ATX Intel Motherboard Could go with a Asus or MSI MB that is NOT desing for OC (Gaming series or even better WS (Workstation) series).
2 * WD BLACK SERIES WD5000BPKX 500GB 7200 RPM 16MB Cache SATA 6.0Gb/s 2.5" Internal Notebook Hard Drive Bare Drive (could be changed for 3.5 inch more performance HDD or SSD).
NZXT Kraken X61 RL-KRX61-01 280mm All-In-One Water / Liquid CPU Cooling Solution (or similar Double rad) You could go with a Single ALL IN ONE water cooler if you need / wish to.
Intel 530 Series SSDSCKGW180A401 M.2 180GB SATA 6Gb/s MLC Internal Solid State Drive (SSD) - OEM This will goes directly on the MB
SAPPHIRE DUAL-X 100314-4L Radeon HD 6970 2GB 256-Bit GDDR5 PCI Express 2.0 x16 Video Card May want something a little beefier if you do 3D.
Hope it will help you.
I wouldn't recommend a liquid cooled system such as the above. For extreme gaming? Yes. For production? No.. You can split the cost and buy yourself a decent rig and a slave render server that would speed up rendering before going OC.
A few tips would be to figure out what you're going to do more, edit photos (max RAM, medium CPU and GPU) or create/edit/process video (max everything).
Adobe products use various parts of your system entirely based on your task. There is no one single answer. You might be editing a 24MP photo but you shift that photo into 16bit color or start batch processing HDR and it can be just as intensive as video and will rely on the speed of components you might underthink, like the speed of your hard disk.
Same for video. If you're just literally editing video and audio with few effects, a $1kish system can handle that at 1080p60. Speaking of, is this for anything stereoscopic?
The 3 components you question have these benefits:
RAM: More images open, larger images won't hit scratch disks, longer video previews, working on videos and images simultaneously, doing complex processing on either or. It's dirt cheap and I can't see any reason for you not to get at least 32GB.
CPU: Faster filters (video/images), faster export processing (video), faster previews (image filters/video), faster multitasking silky smoothness and interface responsiveness. Quad core or better, Intel, almost no exceptions.
GPU: Everything the CPU does with potentially real-time export and playback of video and images with effects (OpenGL cards). The GPU is immensely faster than a CPU at the number of gigaflops it can handle but it only processes very specific data. Never use onboard video, ever!
It's hard to tell you where your specific needs are best handled without in-depth information on what you do with photos and videos and exactly how much processing you do overall. A big batch of that cash will be eaten up by a monitor if you don't already have that.
....no !!!....don't do it !!!!
If you are planning on using the current and future versions of PPro, AND are planning to edit in a professional, or, even advanced amateur manner....you MUST be aware of the following :
1. FORGET AMD....PPro is GEARED to run with CUDA acceleration...this is currently provided by NVidia GPUS ONLY !!!
2. CONSIDER the brand new Haswell E CPU and its new X99 chipset. This new platform will help meet the future...which includes 4K and beyond. TEN onboard SATAIII ports will provide better high speed storage options and the new quad channel memory will improve performance....look FORWARD !!!... not BACKWARD with old motherboards and chipsets. Look at Eric's postings on his recent test results for the Haswell E. Even the mid level and base level CPU perform very well....especially considering their cheaper price.
3. IF using any mechanical HDDs in a desktop, do NOT use the 2.5" notebook hard drives...they are SLOWER than the larger destop drives. HOWEVER, if it is at all affordable, you would do BETTER using Samsung 850 Pro SSDs,or, Crucial M550 SSDs for your storage. The higher speed and low latency adds up to WAY better performance.
1080p60 video is going to require gobs of space. In a $2k budget, there's no room for a $700 1TB SSD when the performance of a couple RAID0 raptors would give excellent performance and quadruple the space, cheaper. Stroked, you could barely feel the difference, except when you simply don't run out of space. And don't get me wrong, I have SSDs in RAID on my video rig, but I had a $5k budget.
I'm aware of the X99 prowess but I cannot sink 1.5K$ JUST on the MB & Processor + the DDR4 price.
It's guides lines I gave, it's a personal build that I change / update with time until I hit completion. The 2.5 inch HDD was a limitation of the previous version of the case (now they added 3.5 inch hdd support).
NB: Z97 is the CURRENT chipset that is not geared toward "Enterprise class" / "Professional workload"
NB2: I'm no near a professional, I need something that will do a lot better then I have with a "reasonable" price tag. Also my Post was geared toward the previous poster price range request and was to serve as guide line.
I would be curious what would you change in my build? Agree that the GFX card need to be switched to Nvidia.
What else would you change and why? (Trying to learn at the same time).
I am editing FINE right now, using an Asus G750 JW laptop that I recently bought " refurbished" for $850...it arrived in a brand new condition...with even some of the plastic on it. I cloned the crappy 5400 rpm drive onto a new 500GB Crucial M500 SSD I bought for $280 on sale. Then, I increased the memory to 24 GB FOR $100. In the second HDD bay, I put another 480GB Crucial SS I had already.
This machine tested VERY WELL...equal to a strong desktop of only 18 months ago....and even better than many desktops currently being tested on the PPBM7 website. You DONT need to " hemorrage" big money to put a good system together.
Nivis Tigridis ~ My main concern with your recommendation is you expressly said it's designed for overclocking. You're using an Intel K processor with a NZXT Kraken liquid cooling setup. All of that is losing precious money for anyone looking for rock solid system performance.
Let me dissuade you from telling me how solid overclocked machines can be. I've been building systems since 80286 (as a job when I was a 14). I had a 286, 386, 486, pentium, etc. I grew up building every single machine by hand having worked in a computer building facility testing and building. I'm as experienced as any enthusiast can be and have overclocked and had SLI video cards since 3DFX invented it. I realize you can build a very solid system, but this is not the system for a person working on 1080p60 or huge 24MP photos. Nobody cares if the OC'd system dies in the middle of battlefield 4 and you just need to reboot. Everyone cares when you're batch processing 24MP photos or are in the middle of editing a huge 1080p60 video and the rig reboots. Stability is absolutely crucial, and overclocking is the antithesis of stability.
So nothing is wrong with your build for a gamer, but it's not designed for a workstation.
JFPhoton ~ The SSD alone is what most people neglect. A system is as fast as its slowest part and while the CPU, FSB, SB, memory and video card can handle gigs per second (or gigaflops for video/SIMD), the hard disk goes unnoticed running on average around 40~70MB/s in your typical laptop. So it's no wonder your system feels perfectly able and comparable to modern systems when loading things. But when it comes to processing power, I'm under the impression your system will show you its true speed when you export a long, heavy CGI video from After Effects. Unless you have a CUDA in there, the raw processing power needed is very high. This is where you wouldn't feel so confident against a basic i7-4970 with even 32GB RAM and a CUDA writing to SATA2. You'd take twice as long, if not longer.
adamissosmrt ~ You're going to need to answer the question. Is video or photographs your main concern? The two have a bit of difference in the proper recommendations. If you do even half video, half photography, I wouldn't recommend spending on a CUDA. I'd recommend using that money in different places. That's the crux of the equation that only you can answer.
TY for took time to reply to me and share your knowledge / point of view on the matter.
And I did mention it to be clear right from the start, since the OP do not seem to be very experience in computer hardware.
Actually standard overclocking or not extreme overclocking is completely stable. This is essentially overclocking the CPU 15% or less. I have also been building systems since the 90's and this has not changed that much over that period of time on average. Some chips released with very high operational temps so OC was not really an option without peltiers and such. The all in one water coolers are fine and give close to the same results as many active fans. The difference is profile and ambient conditions. The water coolers handle higher ambient temp conditions better. The CPU has a large impact on GPU acceleration in media content applications because it has to decode data and create the data buffers for the GPU to do its processing. Then it has to encode the data after. There also needs to be enough ram to handle the GPU acceleration efficiently or the GPU will sit idle far more or even cause errors on exports with many frame resolutions.
I wouldn't recommend a liquid cooled system such as the above
That particular one or any?
Corsair Hydro coolers do not give better cooling performance than air coolers from Cooler Master or Noctua, nor do they run quieter. They are only more expensive. Don't forget that air coolers in push-pull configuration have two fans, that are replaced by two fans on the radiator when using a Corsair Hydro. Unfortunately, the Corsair fans are pretty noisy, thereby defeating the intention of water cooling to make things more silent. Effectively, one replaces two fans with an average noise level of 25 dB with two radiator fans with 35 dB. If the cooling performance was better, that extra noise might be overlooked, but that is not the case. Up to around 4.8 GHz overclock, the performance difference between air coolers and Corsair Hydro is negligent. So, if you opt for Corsair Hydro coolers, you don't get better cooling, you do get more noise, more power consumption (because of the waterpump) and you spend a lot more.
Actually standard overclocking or not extreme overclocking is completely stable.
Yes, which is why an intel CPU will go turbo when necessary. It's already overclocking itself. The extra 15% you can gain can be achieved safely on air alone, while saving precious money. And in that 15% I entirely mean (based on BIOS) using the built-in overclocking drop-down or manually over volting where needed while increasing the FSB, not the built-in turbo.
If the OP even decides to OC, it won't be to water cooling levels, unless they like random reboots and potential corruption. I don't believe it's worth the hassle for such a small performance gain, added cost and the potential for there to be a water hazard in your case at all times (no matter how much they've matured, it's all from China).
Peru Bob wrote:
That particular one or any?
Any. Water cooling has matured greatly as it has been around for quite a long time. It's not that I'm as concerned with it leaking as I am with money being spent efficiently. $1,500 desired but $2,000 max leaves little wiggle room here. We don't even know if it needs to be a notebook so this conversation could be moot.
As for a CPU in this mix we already need to target something as fast and cheap as possible. I think the Intel i7 5820K is the only CPU to consider since it's $389 and performance is similar to many $500 and up chips. Instead of spending $140 on liquid cooling you can get a well vetted 120MM ZALMAN (LGA 2011) for a mere $42, saving $100. It'll still let you get the ~15% OC if you really want to tweak it.
Problem is, even there we're already up to $431 just for a CPU/HSF. That's nearly 1/3 the budget. We still have important parts left, like 32GB RAM, a CUDA video card, quality PSU, case with adequate cooling, fast/large HDs, etc etc. Somehow all of that needs to come below $1,100 to hit the ideal mark (not really possible without significant nerfing of "something") but possible in the 2K range. And consider that budget may need to include a display/keyboard/mouse/etc...
Bottom line is a real powerhouse system costs quite a bit. There's no room for frills.
Turbo is thread based ie it only turbo's to a speed based on cores used. OC turbo's all cores. Big difference there in media content applications.
Yep I edited the original to add that sentence in there. The air alone on the HSF I linked will easily achieve 15% (which can cheaply be slightly upgraded to 2 fans and still be much cheaper than liquid).
One thing to keep in mind is when the system goes turbo, cores are reduced as necessary. Haswell increases 100mhz per bump until achieving the highest turbo using a single core. At that point you're no longer receiving the benefits on all cores and the overclock is minimalized. This is under super duty load, which is exactly what editing media will do.
The only unadvisable way to go to make it worth it is disable turbo and overclock the crud out of the base 4 cores. At that point you should be using liquid cooling and hope to bump each 2.0GHz base core up to 2.5-2.7 max. But we're in dangerous territory, degrading the lifetime usage of the rig as well as inviting instability while increasing cost.
The turbo does not limit the cores used when overclocking. The overclock sets all cores to max ratio regardless of the cores used. Turbo only changes the core usage when your not overclocking.
I've continued my discussion with Sinious in Private message and something that came out is that:
IF OC is not done correctly you will NOT have a stable system and OC = $$$.
I've pass a lot of time on my current OC rig and the other before, I know the performance boost you get but I've also paid a good Prenium to achieve it correctly and Stable.
True a GOOD STABLE OC will benefit anything that is muti threaded and CPU intensive, BUT for the initial price of a WHOLE system it can impact a lot the budget.
I've must put AT LEST 1.5 to 2K ONLY on MB, CPU, COOLER and RAM. That's all ready the budget of the OP... IS OC a option in the FUTURE yes, it can be.
When you use your software is your CPU used at 90%+ or not? Not then why OC? The boost will be marginal. Maybe clearing 1-2 bottles necks will help more for the same "price".
Yes then GO AHEAD that will help a lot more.
That's what I think, OC is there if you need it, but like Sinious said; is it WORTH the price tag... Only YOU can decide.
GPU acceleration has changed the landscape. GPU acceleration requires the CPU to process the data first before the GPU can do it's job. Then the CPU also has to process the final data. The load on the GPU at any point in time along with the performance of the ram in transiting data back and forth between them is decided far more by GHz once the amount of threads are reached for ideal decoding/encoding. Another aspect to this is processing latency. The greater number of core chips can more than handle the processing load of the codecs over time but cannot do so in the time segment required by seamless playback. This is why you don't want CPU's below 2.6GHz for media content applications. Yes the OC addes a higher performance ceiling which equates to greater complexity realtime but it also add's performance in general to the entire pipeline for GPU acceleration regardless of the load put on the CPU. This is why it's important to get a CPU that can handle and push the GPU and vice versus. Otherwise the performance of 1 is idle while waiting on the other to handle it's processing load. 1 last aspect to this is frame sizes. 4K frames have far more data per frame than 1080. Yet the CPU still has to process the same amount of frames in the same period of time. This is where clock speed is extremely key. If you are working with 4K frames then cores and GHz are very important for playback.
The turbo does not limit the cores used when overclocking. The overclock sets all cores to max ratio regardless of the cores used. Turbo only changes the core usage when your not overclocking.
Turbo Boost does exactly that, reduces the number of cores to stay within proper thermal limits (TDP).
However yes, you can have all cores running at max turbo speed. It's essentially overclocking because the max power draw and heat are going to be in use 100% of the time. It's a pretty big waste of electricity and heat management.
To do that you'd need to disable all throttling and power savings (Turbo Boost, SpeedStep and toggle Multi-Core Enhancement). This essentially sets your CPU up as a regular old CPU from the good old days. Then you clock up. Almost 100% of the time, without serious cooling, you can't hit turbo on all cores without TDP interfering with a reboot/fail. This is just Plain Jane overclocking as the cores were never meant to run this hot. If you're lucky you'll only need to reduce each core by a couple or few hundred MHz (3.5 or 3.7 for a 3.9, etc) to attempt to run stable. I'd never do this on air cooling however.
As Nivis said above, for an OP asking about a build, the experience level we're working with here doesn't fit this conversation anymore. While it's not a derail for anyone looking for information or potential options for an editing rig, I don't feel it's headed in the OPs desired direction anymore and is nearing a derail.
Overclocking aside, the OP needs to respond.
Actually Overclocking now is through Turbo. It's just setting the ratio's manually for turbo and assigning the power settings. No you do not disable turbo but you can speedstep on some boards. Some boards have Speedstep tied to turbo so if you disable 1 then you disable turbo as well. Intel moved to the Turbo OC model back with I7 Gen 2. I am sure they intended that for better stability and far less CPU's getting sent back for warranty after bad overclocks. There is no slowing down of cores or GHz ie CPU throttling C-States unless the CPU temp exceed 90C on these boards. That is the throttle point used by most board manufacturers now days. As long as the heat of the CPU does not exceed that point you can clock every chip Intel ships that is not locked like the Xeons to their turbo speed on all cores and that is overclocking. They will run at that speed without exceeding 60 to 65C at all with standard CPU cooling for socket 2011. Air cooling works fine for this including the Noctua fans. The NH9UB will easily handle that but wont handle the Overclocks that get into the 4.4Ghz range. The better Noctua will handle the 4.4Ghz on down.
4K frames have far more data per frame than 1080
4 times the data. 4k resolution is 4 1920x1080 frames in one side by side, top and bottom, yeilding a resolution of 3840x2160 for the record. However the OP said immediately that they were not interested in 4K video.
Yes, a proper CPU selection needs to be made to ensure the GPU is fed but I think you're overestimating what it's doing. With CUDA acceleration in output in Premiere Pro, the CPU is acting much more like a south bridge. It's in data transfer mode, handing the word to the CUDA cores to actually be performed and displayed (directly) after the work is complete. Every effect in the (After Effects and) Premiere Pro panel is accelerated. AME also uses CUDA to accelerate scaling, timecode filters, pixel format conversions, deinterlacing and aspect ratio changes.
As for realtime preview, just watch the video of the Lumetri Deep Color Engine playback in realtime using CUDA versus dual Xeons (16 cores). The link is just above the benchmarks. Per the chart, the K5000 is over 10x faster than pure CPU and I'll spare you the search, that's at a 256bit memory interface and 1536 CUDA cores. The K6000 has a 384bit memory interface and 2880 CUDA cores with 12GB GDDR5 haha. That's insane and yeilds over 20x faster speed.
This isn't over burdeoning the CPU, this is literally what the OP wants. This is the CPU handing off work to be completed and displayed on the GPU, blitted directly after computation to a buffer. It helps a ton in previews and interfaces as much as encoding/decoding/transcoding. Just look up the CUDA cores for each card in the benchmarks to know how much it took for that performance increase.
After this thread I went ahead and purchased a GTX 980 with a 256bit memory interface and 2048 CUDA cores. I'll be happy to report back after I replace one of the video rigs which has a Radeon HD6870 currently only slightly accelerating the interface but obviously zero CUDA acceleration. The CPU on the system is always maxed out when rendering so the CUDA acceleration and CPU consumption comparison will tell you exactly how much work is offloaded from the CPU, rather than bottlenecked by it.
AE does not use CUDA acceleration. AE rarely uses the GPU at all. AE had Ray Tracer which is completely different and rarely used. Premiere Pro uses the CUDA hardware acceleration for many effects, scaling, interpolation for the most part. I am not overestimating what the GPU acceleration is doing. I am describing how the GPU acceleration works because I actually read the CUDA white papers. I know how it works. And in the media content industry right now the CPU's are handling the decoding/encoding for most part and some effects along with all of the memory management ie including ram buffers used for the GPU data. That is a very significant load. I have benchmarks over the last 4 generations of Intel chips that show the progression of CPU load versus GPU load when GPU acceleration was available. For example until Haswell E a single CPU could not regularly push a Titan card over 60% or so. Now with Haswell E 8 Core and Xeon E5 2697V3 Xeon 14 core we are finally starting to see the Titan load at 90%+ constantly and while the Xeon CPU is still at 50% load or less. This has not been the case at all in previous gen platforms. You can also reference this by the fact that the lower GHz CPU's were not loading the GPU's as high with the same exports. Once again the benchmarks showed this if it was monitored. The export times were also effected by this as well so the export times reflected the same data.
There are benchmarks all over the place that show the GPU's are significantly increasing the performance over frame aspects that take the CPU a long time to do. That doesn't however mean the CPU's have less work in that same time segment. It just means the CPU's are more focused on the work they are doing in that same time segment. To give you an example. I playback my R3D 5K sequence the CPU load is 100% whether I am on software or hardware MPE mode with the 5820 CPU. The GPU load is obviously vastly different between them. I put in the 5930K CPU and the load averages 60 to 70% or so with the GPU acceleration on. However once again the playback is much smoother at full resolution preview and the GPU load is vastly higher. That is due to the fact the CPU still has allot of work to do besides what the GPU is processing currently and that is obviously with the GPU accelerated Debayering as well of 2014.
Got to love the benchmarks where they compare with Quadro cards but not the Geforce cards. The Titan Black and the K6000 are the same card minus some extra vram on the K6000 with regards to specs.
No the CPU is not always maxed out rendering anymore and hasn't been since I7 Gen 3 ie Ivy - E. Haswell E is Gen 4. Right now the only application I saw max out the E5 2697 V3 14 Core consistently was AE. Often times Premiere did not. GPU-Z will tell you what the current load and recent history load on the GPU is during that time. You can use that in comparison to the Windows Task manager performance meter and CPU load to monitor the parallels between them. You can also look at benchmarks such as the PPBM series and see the same GPU's in different systems with different CPU's perform far different with regards to the GPU acceleration. Why do you think that is if the CPU's are not the primary factor involved?
BTW I put the information about 4K out there for others who read the forums.
Adobe After Effects does use CUDA, here's the relevant info and a complete video on how to unlock/configure CUDA performance all the way down to the interface:
I do think the CPU is involved. It's the master, the GPU is just a slave (bandwidth to be abused at will). The CPUs role simply changes when it doesn't need to be the one processing the frames. It still is managing multitasking, tons of I/O and anything any particular GPU can't accelerate.
Again, I have no CUDA now and I have lots of 1080p24 work I'm doing currently. I did these jobs myself on the same workstation. I will be intimately aware of even the slightest bump in interface, scrub, preview and export speed. I will report back on real world non-CUDA vs CUDA. This is AE CC 2014 hefty work with 3rd party plugins even (which raytrace).
I don't expect a first generation i7 to act like the 4th generation and I do read the release notes (not the whitepapers, I don't care that much haha) for the bullet points on optimizations. The key in 4th gen that's most relevant is AVX 2.0, doubling the SIMD interface to 256bit and doubling floating point performance. That's a worthy reason to upgrade alone.
I'm not entirely clear on why it's news that a faster, later generation processor will keep a faster GPU fed. Generations have largely maintained the same speed range but like AVX 2.0, the efficiency improvements are helping the processor simply do more work in the same time frame. So of course moving to a faster processor is going to tip the scale of "who's the bottleneck" back at the GPU. But the work the GPU is doing for the CPU is where I think we disagree. You're saying the CPU is still doing the work and I think it's pretty clear that the CPU is offloading that work to the GPU. The CPU does not do what it used to do in the presence of CUDA acceleration.
What would be nice, when you really need to get the work done, is much like a network rendering farm, the CPU should take on work it's offloading to the GPU to even further speed up the rendering. It should be an option because I myself don't won't to purchase 2 machines at home for video. I want enough 'spare' CPU left over to render in the background and multitask otherwise. So I don't want the CPU pegged out at any point in time. But only being at 50%? That's a huge waste to me.
CUDA is almost entirely irrelevant to After Effects. Only one thing is accelerated by CUDA: the ray-traced 3D renderer.
Todd Kopriva, Adobe Systems Incorporated
After Effects quality engineering
I should have phrased it better as AE Ray tracer acceleration is vastly different than Premiere MPE GPU acceleration. IE AE rarely uses Cuda while Premiere uses it for processes all of the time.
I use that a ton. I use the built-in and also VideoCopilot's Element 3D in almost every single video. It's PAINFUL to scrub the timeline, let alone export. If that performs all the calculations the CPU is doing for me now, that will reduce my rendering time substantially. I have literally thousands of 3d particle effects and lots of lights on many complex textured objects.
That said, not everyone uses the 3D features of AE so in fairness, those of you just editing videos won't get any of that.
However, Premiere Pro and AME state they do make use of CUDA when previewing and exporting with huge speed increases. That's where I feel the OP is largely going to be. Rather than generating CGI, just using Premiere Pro to edit amateur videos and AME to export the final product. A worthy target to desire CUDA acceleration.
Keep in mind, products get regular upgrades. Knowing the potential of the technology, do you see them doing anything but embracing it everywhere it can possibly be used? When you compare the raw power of the CPU and GPU, it's no doubt anyone would love to tap the GPU. But the work must be within the limits of the beast and a GPU is really only good for very specific things. One thing is for sure though, video is one of those things, and it's not going anywhere.
Well the last I heard AE is moving away from Ray Tracer and leaving that aspect to C4D. C4D currently doesn't have GPU acceleration built in. I heard rumors of work on it but that's it. However those who use C4D can get Octane which is the GPU acceleration renderer for C4D. Then you do have the GPU acceleration and some really nice features.
I use a lot of 3D in AE so it makes sense for me. That's my boat.
I also use C4D because I create stereoscopic (3D) content and AE with the default 3d camera rig is only capable of half (Horizontal Top and Bottom) or quarter resolution (Half Side By Side) 3D output. Essentially it outputs a 1920x1080 picture with the left eye and right eye data separated into 2 squished frames sitting side by side or on top and bottom of each other. Because each eye is only getting 1920x540 or 960x540 pixels of resolution to fit in a single 1920x1080 frame, it's half resolution 3D. In order to get actual (frame packed) true 3D for both eyes, you need to have Cinema 4D and some helper plugins to integrate them.
Blu-rays come with full data for each eye, called FHD3D. It's 2 full 1920x1080 images stacked on top of each other with 45 pixels of black separating them. These jumbo frames require HDMI 1.4 to serve them at high speed (60FPS+) and are 1920x2205 resolution. AE for whatever reason has no export option for that which is pretty stupid. Neither does Premiere. So I'm stuck rendering final 3D content in Cinema 4D so people with active displays get the full resolution picture to each eye, not half or quarter.
I use 3DS Max with Vray which uses CUDA cores as well.
Overall I'm hoping it's pretty noticeable. And I really don't think Adobe is always going to depend on other products to handle shortcomings. They have to deal with things like jumbo frames now because ultimately they're still a lot smaller than 4K resolution, and everyone is heading there, like it or not. We need the CUDA and OpenGL/CL speed.
Open GL doesn't really provide any acceleration anymore. What was utilized in the past has far been surpassed by CUDA and Open CL.
OpenGL spec and coding aren't going anywhere. Most quality games are developed in it for the multi-platform capabilities (same code runs with minor tweaks on Windows, Linux, OSX, etc. Much like flash, there aren't many serious alternatives that are multi-platform. DirectX is the biggest threat and as you know, that's squarely Windows-only.
But I know you mean in the era of tons of proprietary rendering engines, general purpose SIMD processing acceleration is desired so you can choose any engine as the end target and benefit from the acceleration. And I agree, people are making development tools based on OpenGL much less these days and CUDA-esque techniques are far more desirable.
As an API Open GL is here to stay. As you guessed I meant the acceleration plugins added for it. Those are really legacy at this point and likely wont be used unless something new is changed in the future. Windows unfortunately is the dominant platform currently so Direct X is definitely pushing Open GL use back some. The time it took for the Open GL standard to catch up to Direct X 11 really didn't help either.
This is a long one, skip through to quoted tests if you like.
I ordered my card last Wednesday and newegg next day hath let me down. It didn't come in until yesterday (day off). I did finally have time today to do some experiments and I've learned a few things that people should know about codecs and CUDA.
CUDA has several versions and a SDK (of course) to use it. At present Adobe Media Encoder CC 2014 and Sorenson Squeeze 9 do utilize CUDA, but do not support the new Nividia NVENC SDK for Kepler+ cores. Essentially a couple years of CUDA core cards aren't usable for acceleration in MP4 H.264 MainConcept, the only codec optimized to use CUDA cores (that anyone typically cares about).
I decided not to put the card on a i7-4xxx video rig, instead choosing to put it on the oldest i5-2500K machine. I plan on upgrading this rig anyhow but I wanted to see how much CUDA could help an old machine feel new again.
You won't be disappointed..
It took quite a bit of digging and one great plugin offered that I will link later along with some 3rd party utilities to actually use Kepler+ CUDA.
Intel i5-2500k Sandy Bridge (No OC)
16GB DDR3 Corsair XMS Vengence 1333
3:20 HD (1920x1080p@24) Quicktime Animation Lossless w/ PCM Audio
Video Test Targets:
MP4 H.264 25000/384k 1920x1080p@24 w/ AAC
MPEG2 25000/384k 1920x1080p@24 w/ PCM/AAC
Exports were concurrent in both encoders, meaning both videos were encoding in different threads at the same time.
24p isn't really a difficult target and the source content was multimedia, which is usually pretty tricky to encode, with lots of fast movement and particles. I'm keeping it there so the i5-2500k doesn't cry for mama. I'm also reading and writing to separate disks to remove as much HD bottlenecking as possible.
Each test was performed 3 times and times are averaged.
Adobe Media Encoder CC 2014 via old Radeon HD 6870:
Mercury Playback Engine Software Only
09/30/2014 10:48:50 AM : Queue Resumed
09/30/2014 10:55:00 AM : Queue Stopped
6min 10sec Total
Mercury Playback Engine GPU Acceleration (OpenCL)
09/30/2014 10:59:03 AM : Queue Resumed
09/30/2014 11:05:13 AM : Queue Stopped
6min 10sec Total
Well, that does literally nothing useful at all. Either that old video card couldn't be utilized or is just unsupported.
From here on, the GTX 980 was installed and no more tests were done on the HD 6870. I just wanted a point of reference.
Adobe Media Encoder CC 2014 via GeForce GTX 980 (2048 CUDA cores):
Mercury Playback Engine Software Only
09/30/2014 01:05:55 PM : Queue Resumed
09/30/2014 01:11:57 PM : Queue Stopped
6min 2sec Total
Mercury Playback Engine GPU Acceleration (CUDA)
09/30/2014 01:14:52 PM : Queue Resumed
09/30/2014 01:20:56 PM : Queue Stopped
6min 4sec Total (???)
What the heck? Ok, reading some I see I need at least 512MB RAM on the video card and a compute ability of 1.1 or higher, only then will "use GPU" be enabled. Well, it's enabled and the driver (latest) is reported clearly:
4GB RAM (in the mem tab) and 2048 with a compute capability of 5.2, so yep, CUDA should be working here. What's the missing ingredient?
MainConcept MP4 H.264.
Currently they do not support Kepler or newer CUDA cores at all. So the video encoders are just shrugging it off, neglecting the abandoned SDK for the new NVENC SDK.
To show the unanimous support missing:
Sorenson Squeeze 9 via GeForce GTX 980 (2048 CUDA cores):
GPU Acceleration Disabled
GPU Acceleration Enabled (Only on MP4 H.265 MainConcept)
6min 44sec (???)
Well that stinks! But alas, with much searching and a long face, I came upon this Adobe forum thread with an excellent solution for Premiere Pro and AME, an export plugin that has been updated to use Kepler+ CUDA cores called NVENC_encode! (Thank you!):
Yes, it is experimental and requires 3 3rd party libraries to help finish the encoding, but it actually works perfectly, AND in parallel! After you place the export plugin in your AME CC 2014 plugins folder and restart, you now have a NVENC_encode target to choose from. This source is not Multiplexed (audio/video combined) so as the author recommends, do download the TsMuxer, MP4Box and NeroAACEncode libraries/apps. His plugin asks for the executables paths to each of them and now not only can you export MP4 H.264 with Kepler+ CUDA cores, you can export MPEG2 as well!
That's pretty amazing.
Before using CUDA, here's what every export always showed, pegged out CPU, almost all my RAM in use and the GPU usage at 0-1%:
Now, use the above supplied plugin and here's the new encoding times.
Adobe Media Encoder CC 2014 via GeForce GTX 980 (2048 CUDA cores) using NVENC Plugin (+ 3RD party apps for mux/audio):
Mercury Playback Engine GPU Acceleration (CUDA)
VIA NVENC 3RD PARTY ENCODER (w/ TSMUX (MP2), MP4BOX (MP4) and NeroAACEnc (MP4 AAC))
09/30/2014 04:06:57 PM : Queue Resumed
09/30/2014 04:08:20 PM : Queue Stopped
1min 23s Total (!!!!!!!!!!)
The quality of the video is, even at close inspection, nearly identical to the software rendered version. If there are differences, nobody will notice them. I realize the Quadro cards are designed to be absolutely identical but even when I go frame by frame on either MPEG2 or MP4 source with a software render, I can't see anything at all that I don't like.
One thing I DO like? The encoding time. Encoding a single file let me use up to 50% of the GTX 980, keeping CPU in the 80% range. I could notice my system was much more responsive during encoding:
Even better? Encoding both targets (MP4 & MPEG2) let me use up to 70% of the GTX 980 and still maintain the same CPU. So even on a palsy i5-2500K with DDR3 (in dual channel) with SATA HDs, I'm getting a huge benefit. Here's encoding 2 at once:
80% range was maintained and even though the videos were concurrently encoding I could totally use my system to do other things. In fact the encoding was so fast I didn't even have time to try to do other things, it completed so fast.
Adobe should really jump on board and take a look at NVENC. The advantages are obvious. Not just in encoding. CUDA cores are general purpose power, not specific to video related tasks. Every application can make use of the insane amount of power available, idling on the video card.
And hey, if you're on a budget like the OP, the certain newer and older cards (Fermi, Tesla, ION) CUDA cores can be used by AME and Premiere right out of the box. Just be sure you're not using Kepler or Maxwell.
I can only imagine what the encoding time is going to be reduced to when I quad channel 32GB DDR4 off SSD to an i7-5xxx Haswell hex core with hyperthreading. I'm going to need a second pair of underwear.
Thanks for the head up, I use a GTX760 (Kepler), I'll look into what you've mentioned to see IF I can really benefit from all this.
As for a new X99 platform test here a vid that mention his previous Z97 encoding time and the new X99 encoding time: Watercooled 5960X Overclocking and Benchmarking - How are the temps? - YouTube
I'm totally for water cooling and overclocking. His setup would be something I have done numerous times back when liquid wasn't nearly as safe as it is now. It was the age when we used CRTs at an average 120hz so you actually could feel the difference between 60fps and 120fps. 60fps was boggy, so I overclocked like crazy.
And I get why his 22% performance increase in rendering is considerable to him. Here's what I wonder however. In the comments I see he's running 3 780s, is that correct? Because that puts him squarely in Kepler+ range and he's not even utilizing it. He's impressed that he can export at 78% of realtime. If he had Premiere Pro and the NVENC_encode plugin for it, on his rig, he would probably export 20x faster. Although there have been many disclaimers that using CUDA cores across multiple cards can cause issues and some are limited to only the CUDA cores on a single card.
Bottom line, he bought an insane water cooling system and extremely expensive CPU and moderate RAM (only 16GB? weird). He pretty much doubled his previous CPU. The CPU I recommended above would be incredibly fast as well but cost 1/3. Since he doubled his scores, let's say he doubled his render speed. Then add 22% to it for the overclock as he said. So he's encoding his videos 122% faster than he was before, sub-realtime.
I just increased my encoding speed by 600%, merely using CUDA cores, on a single GTX 980. I think the CUDA cores, given their inexpensive nature, are the way the OP should go.
If I get more time later in the week I'll try to put together a pretty aggressive Premiere Pro project with plenty of effects and give the timing with CUDA on and off. While encoding is clearly insanely better on CUDA, it remains to be seen how much faster actual rendering will be.
Just to be clear most of his Hardware is was all ready his for years so ... Cost wise... and he tell you that it's expensive true but for him it's worth it.
Also he using VEGAS not PREMIERE, it could that the CUDA implementation is different.
I know my post above was too long but it took a long time lurking in the dark to realize what the issue was.
Specifically for encoding in AME, the issue was MainConcept H.264 being the CUDA-friendly encode target. The, PRE-Kepler encoder. Now that's the missing detail here. MainConcept H.264 is a staple target that ALL video editing apps can target. It makes no difference if it's Vegas or Premiere, you are using the very same encoder.
What the plugin I linked above does is give you a H.264 (NON-MainConcept) encoder that can use Kepler+ CUDA cores by implementing NVENC SDK. MainConcept told me they do NOT plan on updating MainConcept H.264 for Kepler+. So it will never render using modern CUDA, in ANY video editing app.
You can just search for Vegas CUDA and see endless people misunderstanding why they upgrade their CUDA and suddenly rendering and performance is terrible and no GPU is used during export from Vegas. They don't know it doesn't support Kepler+. That said, neither does Premiere or AME, without the plugin I linked.
My i5-2500k GTX 980 can probably beat his 8 core overclocked Haswell on export. When I upgrade to the hex core Haswell, I will annhiliate him with CUDA, guaranteed.