Are Realtime HD Previews Possible With An NVIDIA QuadroFX 5600?

[Jimmy Brunger]

Hi Brendan,

A beefy GFX card really won’t help you much in AE I’m afraid. Open GL previews are usually not to be trusted for accuracy either. I may be wrong, but AFAIK it is the procs & RAM (and fast RAID drives to a point) that gives you the previews, not the GFX card. If you are working with lots of 3D lights, etc then OpenGL may give you fast response with certain cards, but not for running streams of video.

I’m not sure what you plan to do in AE, but if it’s just running several streams at once (eg: Picture in Picture or similar) then use something like Premiere or Final Cut Pro and make sure you have fast striped drives and at least 4 procs for full 1080 HD. There are hardware accelerated products out there, most notably ‘Axio HD’ by Matrox which runs with Premiere – that will give you more RT HD streams than most other edit systems on the market, but it’s not cheap.

Those big Quadro cards are mainly useful for 3D apps and colour correction systems like Assimilate Scratch or similar, or as part of a Smoke/Flame system. After Effects really doesn’t take advantage of them at all.

Depends what you are wanting to do, as to what system/software you should be using?

Hope some of that helps.

*Production Studio Premium CS2 / *Combustion 3 / Mocha v1
————————————-
Win XP Pro SP2 / Intel P4 3GHz / 2GB RAM / GeForce FX5200 / DeckLink Pro / Roland DS-5 monitors / Sony BVM-20G1E / DVS SDI Clipstation / Wacom Intuos 3 A4 / 110GB boot/80GB media/600GB RAID-0

[Brendan Blair]

Hi Jimmy,

Yes, what you have said does make sense now that I think of it. I actually plan to do compositing and rotoscope work with full motion images so Premiere (or most other video editors) wouldn’t have this kind of robust functionality to meet my needs.

However, as you mentioned, it would really be about the RAM and the processing power of the computer itself, and not the actual graphics card, that would make all the difference. I will therefore stay focused on that. I just thought that maybe I was missing something by not looking at the video card in all of this.

Thanks!

Brendan

[Jimmy Brunger]

If you’re doing much intensive rotoscoping it might be worth you looking at Motor or Mocha from Imagineer Systems – which use a unique tracker to help you roto. There was a freelancer 90% deal on the software earlier this month, but might be worth checking to see if they’ll still honour it? A decent GFX card will help with interactivity in these apps if you’re using 1080+ plates, but a 5600 is still overkill. Get an FX1500 maybe if you can afford it and if you’re doing alot of 3D work, otherwise use Win XP64 or Vista64 or OSX and fill up with RAM. Also, if you buy Nucleo (or use AE CS3) you can render alot faster if you have 1-2GB RAM per core.

*Production Studio Premium CS2 / *Combustion 3 / Mocha v1
————————————-
Win XP Pro SP2 / Intel P4 3GHz / 2GB RAM / GeForce FX5200 / DeckLink Pro / Roland DS-5 monitors / Sony BVM-20G1E / DVS SDI Clipstation / Wacom Intuos 3 A4 / 110GB boot/80GB media/600GB RAID-0

[Brendan Blair]

Thanks for the information Jimmy.

From what I understand with the way a 64 bit OS works, however, I don’t think it would actually be possible to supply CS3 with 1-2 GB of RAM per core, DIRECTLY (when there are more than 2 cores – actually 3 but you can only ever have multiples of 2 cores). This is because even under a 64 bit OS CS3 can still only see up to the 4GB addressing limitation of a 32 bit app (and of course as we know, CS3 itself can really only see up to 3GB of that 4GB address space). As a result, it would theoretically supply each core with only up to 384 MB of data (3072 MB RAM divided by 8 cores).

My own guess about how to improve the amount of memory allocated per core might be to either create a large RAM disk in the 64 bit OS which CS3 could then leverage for a source drive (it’s not quite the same as directly dedicating a certain amount of memory per core but you have to admit that access times wouldn’t exactly be much slower either). Another way to handle it (or even in addition to the above), would be to locate the CS3 cache directory in a RAM disk as well (though I’ve never actually checked to see just how large that cache directory can get).

Anyway, these have just been my thoughts after reading this interesting article:

https://blogs.adobe.com/scottbyer/2006/12/64_bitswhen.html

Brendan

[Jimmy Brunger]

Brendan,

you assumptions towards AE as an app are correct…it will only ever see 3 or 4GB RAM even in a 64bit OS. However the extra RAM is basically for the extra ‘instances’ of AE that are set off when working with Nucleo or CS3’s multiprocessing mode. Each core works with a new instance of AE (essentially upto 4GB RAM per instance) This speeds up previews & renders significantly. 8 cores and 16GB RAM will be a beast of a machine. You will need a 64bit OS for this kind of workflow of course.

I am still on a cr@ppy single core P4 and XP32, but we only do SD 30 sec commericals on the whole. Our work is starting to get more complex of late though and with HD on the horizon it looks like 8cores and Vista/XP64 is the way to go. Mac is an option, but AE is limited to 3GB per core with mac and we use Lightwave which doesn’t have a mac version…That said I’m setting up a Mac system at home soon.

*Production Studio Premium CS2 / *Combustion 3 / Mocha v1
————————————-
Win XP Pro SP2 / Intel P4 3GHz / 2GB RAM / GeForce FX5200 / DeckLink Pro / Roland DS-5 monitors / Sony BVM-20G1E / DVS SDI Clipstation / Wacom Intuos 3 A4 / 110GB boot/80GB media/600GB RAID-0

[Kevin Camp]

with mac (probably pc too) you can hand 2gb of ram to each core for rendering, if multiprocessing is enabled in ae (cs3). the way multiprocessing gets around the 32bit ram addressing is by running a separate render engine for each core. each render engine is then allowed it’s own ram allocation, so in effect the each core gets it own ram allocation too.

soon (rumors are now mid january), cs3 will handle the p2/mxf files directly, so you won’t need to convert them to avi.

however, to get as fast of processing as you can with ae, you’ll want to convert your dvcprohd to something that does not compress with temporal/interframe compression.

uncompressed or none would be best, particularly if you can use a 16pbc codec. lossless animation is pretty good, although it is 8bpc. image sequences are also very good (tif, targa, png. etc). if you had a mac and fcp prores would be decent, it will maintain the 4:2:2 color space, but only uses intraframe compression (although, i’m not sure about decompression load). avid has a very similar hd codec you can download from them. photojpeg at high quality settings may also work well too.

if you use a codec that produces large files (like uncompressed), you will need to have a fast disk array, one capable of data rate fast enough to play back at least one stream of uncompressed hd in realtime. a 4 drive sata2 striped (raid0) drive array may achieve that, and 8 drive array should do that with out a problem.

Kevin Camp
Designer – KCPQ, KMYQ & KRCW

[Brendan Blair]

Ahhh, I see! I obviously didn’t quite understand the way CS3 would launch separate processes to accomplish the rendering…

Thanks!

Brendan

[Brendan Blair]

Thanks for such a detailed response Kevin!

My only question when thinking about what you and Jimmy have been saying, is that if the RAM were increased from 16 GB to 32 GB (I realize this really starts getting expensive though), would there be a further performance increase that is even worthwhile, or, once each core and rendering process has a certain amount of RAM then perhaps no further benefits can even be gained? I am wondering where that sweet spot might be, if there is one. So in other words, would there continue to be a gain going from 16 GB of RAM to 32 GB (even if you can’t quite use all of it amongst the cores), or would this be overkill because 16 GB would actually be about the best performance you could get?

An example would be creating larger and larger buffers for DVD burners. Eventually there just isn’t a point in increasing the buffer any further because no more performance gains will actually be realized.

I have only seen tests done with 16 GB of RAM on the Mac, but no one seems to have tried 32 GB (the Mac can now be further expanded to this, otherwise XP64 bit would do).

I just mention this because the more cores that are added, then the smaller the memory speed becomes because the “pipe” must now feed 8 cores, rather than 4, or just 2. As a result, having a lot of RAM would create a buffer that might help alleviate this memory speed bottleneck, but to what degree?

Any ideas? I’m looking to purchase a PC or Mac system near the beginning of February after hearing the results of Steve Job’s keynote address. I know the FSB should be going to at least 1600 MHz, and the corresponding RAM speed to 800 MHz, (note that Dell’s workstations can already be configured for this – even though their website “configurator” is actually messed up at the moment), but I’ll also wait to see what Apple comes out with as well.

Brendan

[Kevin Camp]

you may start to see diminishing returns as you max out the ram per core. you may see that earlier too if you are not working in a resolution that will push your system (sd won’t tax your ram like hd or 2k projects would).

i only know of one guy with 32gb of ram in a macpro. i got the feeling he wasn’t as impressed as he had hoped with cs3’s multiprocessing and 32gb of ram. he has mentioned that nucleo pro did make things work better, and recommends it over cs3’s mp renderer. there are many advantages to nucleo pro, like background ram previewing and rendering. i’ve also heard that it seems to handle ram better, like it is somehow more organized or optimizes use somehow.

it’s true that the pipe feeding the ram is only so big, and only so much data can flow thru it, but if your pool of ram is bigger, that greater capacity is used to better feed the processors with a steadier stream of data.

Kevin Camp
Designer – KCPQ, KMYQ & KRCW

[Brendan Blair]

Thanks for the real-world information you obtained.

In numerous tests I’ve done it seems to me that the best way to utilize more RAM is to work with uncompressed images (that is, pre-render the video tracks) and put them in a RAM disk. Of course, this only works if the total length of all tracks is very short since the availabilty of RAM is limited (and you would still need a lot of it). Otherwise, a RAID array would be the next best thing for performance of course, and then tracks could be as long as needed.

The reason I mention using a RAM disk is that if the source files are already compressed in some format then the speed of a RAM disk compared to a RAID array, or even just a single drive would be largely irrelevant. This, of course, is because even a single hard disk could keep up with a few streams of this type of data. However, the overall rendering process would be significantly sped up if actual uncompressed frames were in a RAM disk where they could obviously load significantly faster than even the fastest RAID array.

So the way I see it, you copy the uncompressed images from a hard drive to a RAM disk just once per session, and then as you spend hours refining the effects work, all rendering happens incredibly fast. This would be a great way to utilize more RAM (or else maximize the RAM per core and use a RAID array as the other alternative like I mentioned above. Of course, if going the route of a RAM disk you would also want to ensure the system is battery backed!)

I realize, though, that this works great for my workflow but many others might be working on much longer video tracks so perhaps it wouldn’t be of benefit. In one such case I was working on a 33 second segment that would take about 10 minutes to render each time. If I had had uncompressed frames in a RAM disk then every time I had made slight alterations to an effect then I could have seen the results so much faster.

So this is why I’ve been considering a lot more RAM. Not just to increase the RAM per core, but to provide a RAM disk for doing rotocope work that requires an entire video track to be re-rendered each time, just to see the results of even a small change. Of course, if this were also combined with the background rendering of Nucleo Pro then the results would be even better still.

Anyway, just my thoughts…

Brendan