AE CS6 11.0.1 CUDA BENCHMARK PROJECT – test your graphics cards!

[Ian Mapleson]

The scaling does drop off somewhat after two cards, but more
still helps, though I found the gain from three 580s to four
580s isn’t that much. I suspect it’s probably not worth having
more than four 580s, ie. for the performance gained vs. extra
power consumption, etc.

Ian.

SGI Guru

[James Osbun]

New result with dual GTX 580s 3GB RAM each.

2 min 52 sec

[Dmitry Kitsov]

3x GTX Titan
Windows 8.1

1:32

[Ian Mapleson]

Excellent!! Now we’re talking! 8) Is that tri-Titan time
done with the Titans running at stock speeds?

Ian.

SGI Guru

[Dmitry Kitsov]

Yes, though one is EVGA factory “Superclocked” at 875. Other two are at 836. I am also running them at PCIE2 not 3 (no mobo support for PCIE 3)

[Ian Mapleson]

Sweet!! 8)

Note that I doubt the PCI 2 vs. 3 config makes any difference
for this particular test.

So, now all we need is a result from someone who has one
or more 780 Tis (should in theory be quicker than Titans).

Ian.

SGI Guru

[Steven Andrus]

Dual Xeon e5-2690 at 2.9ghz (speed step kicks it up to 3.2 or something I think but I didnt see the render really hit the cpus at all so i think it stayed at 2.9)
64gb of 1600 ram
windows 8.1
Single asus titan 6gb
ran the benchmark with a browser open on another monitor driven by the same gpu and had a few other apps open and system monitor open. I also ran it with one display and separately with caps lock on, same time every time. Honestly we should be running this at 4k though. I’ll look into converting the file into 4k and link it if I do.

Render time: 3:13

[Ian Mapleson]

Steven Andrus writes:
> Dual Xeon e5-2690 at 2.9ghz (speed step kicks it up to 3.2 or something

Yeah, does that on my Dell T7500, usually stays at one BIN above the baseline.

> I think but I didnt see the render really hit the cpus at all so i think

It’ll barely touch the CPUs at all, being a CUDA test.

Real-world datasets can hammer the CPU(s) aswell at times, but not this
test, it’s pretty simple & repetitive.

> Single asus titan 6gb

Just FYI, a 780Ti should be quicker.

> Honestly we should be running this at 4k though. I’ll look into
> converting the file into 4k and link it if I do.

Not sure it’s worth doing. All it would do is quadruple all the running
times, except systems where RAM on the GPU suddenly becomes an issue and
they’d be even slower, even though such cards aren’t fit for 4K anyway.
Running it at 4K wouldn’t reveal anything new.

This test is interesting as a means of testing one narrow performance
aspect of AE (namely CUDA on a small dataset that doesn’t hit CPUs, RAM
or I/O), but for me it’s thrown up questions for which I can’t find
answers, eg. is it possible to force AE to use multiple GPUs round-robin
for frame rendering? AFAIK atm the app always tries to use all available
GPUs at the same time for every frame, which often scales very badly
indeed (extremely badly in some cases). Look at my multi-580 results for
a good example (see earlier posts): with more than 2 cards, the
exploitation percentage of each GPU drops off sharply, so four 580s is
barely any better than three. Performance would be much better if the
frames could be rendered 1-frame-per-GPU, so with 4 GPUs the first GPU
would render frames 1, 5, 9, etc. I can’t see any setting for this in the
Settings panel though.

The other question is, does AE ever make use of 64bit CUDA? That’s the
only real advantage of Titan. Are you running the Titan in 64bit mode?
If not, try it in 64bit mode, see what happens, though that would
probably only reveal whether this particular tests gains from 64bit mode,
not whether AE uses it in general, and if so then to what degree. If AE
doesn’t need 64bit CUDA, then (except for the lack of ECC RAM) the best
value CUDA card atm for AE is the 780 Ti, unless somehow one is running
up again the card’s 3GB RAM limit. Most likely though some vendors will
eventually release 6GB 780Ti models.

The final question is whether AE would benefit from the full speed PCIe
return paths found in Tesla cards, and the better GPU cache structure &
other additional features. Is AE even coded to make use of these? Who
knows – there’s nothing on the Adobte site about this.

A friend of mine is working on a more real-world dataset, a 30 second
animation which atm takes about 2 hours with a couple of 580s. It hammers
the whole system and so is a good general test, including system stability.
Not ready yet though.

Ian.

SGI Guru

[Tyler Tometich]

Mac Pro (MacPro5,1)
2 x 6-Core Intel Xeon @ 2400 MHz
32 GB – DDR3 ECC @ 1333 MHz
NVIDIA GeForce GTX 680 2GB
OS X 10.9.1

RENDER TIME – 6 Minutes 46 Seconds

*dIGITAL cREATIONS*
-Mac Pro (MacPro5,1)
-2 x 6-Core Intel Xeon @ 2400 MHz
-32 GB – DDR3 ECC @ 1333 MHz
-NVIDIA GeForce GTX 680
-Canon 5d Mark III
-Adobe Creative Suite CC

[Cyra Emery]

I must be doing something wrong as it finishes in 17 seconds on my hackintosh.

2x GTX Titan (EVGA Superclocked), E5-2695 V2 2.5GHz 12-core (5% OC on it), 64GBs 1600Mhz ECC RAM with an external SAS Raid.

mavericks 10.9.2 (1st beta).