[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

[Ian Mapleson]

Roberto Tafuro writes:
> kill myself (in the “night-render” world you know,the important thing is
> to make all right and don’t need to render again!), …

The people I help want both. 😀 Fast & reliable.

> … but i don’t understand
> why a GTX680 with 2GB renders in 6.11 min and a GTX680 with 4GB renders
> in 7.4 min. …

Probably the latency is different because of the way the RAM is setup using
different ICs. And are you sure both 680s have the same core/RAM/shader clocks?

> … So a GTX 680 will render in approx the same time on every system.
> Is it correct? …

More or less, yes. I’ll be testing this to the extreme soon using an i3
550 system and four 580s, but so far I’ve not seen any excessive
differences with respect to the test platform being used.

> … PCI 2 or PCI 3 as i understand come in place where you use
> SLI so a PCI 2 doesn’t make a difference on a single card.

It might also matter when using multiple GPUs even without SLI, though
I was told the return path for gamer cards is only x1 speed anyway so
it doesn’t matter, but yes in this context I doubt it matters.

> I’ve tried to boost the card in GPU tweak,but i’m not able to oc a card
> and the fear of screw something is too much,so i prefer to work with
> standard settings…

I wouldn’t use any of the 600 series cards for CUDA. I realised this was
the case when the first reviews were published. Plus, back then, none of
the 600 cards were quicker than my two 560Tis SLI so I didn’t think it
worthwhile even for normal 3D gaming performance. Now I have two 580s
SLI which is faster than a 780 for gaming and waaaay faster for CUDA;
hence, my CUDA research machine has four 3GB 580s, and I’ve just obtained
another MSI 3GB 580 L.E. which should oc to 1GHz+ (I was able to give it
a quick run at 950MHz no problem; see https://www.3dmark.com/3dm11/7390717).

> I’ll stay with this values and be happy with the card,but the 50sec
> differences from the result i’ve seen here with another 680 is really
> strange to me…

Hard to know for sure, but check those clocks with GPU-Z, and as I say
the latency through the RAM could be a factor.

> Maybe there is a way to use better the GTX680?

Yes; for CUDA, sell the 680 and use the money to buy a 3GB 580. 😀

> Thank you again for your time and patience and sorry for my bad english!

You’re kidding right? You have better spelling, punctuation, grammar,
etc. than most Brits. 😀

Cheers! 🙂

Ian.

SGI Guru

[Ian Mapleson]

That’s why I fitted my system with four 580 3GB cards. 😀

For 600/700 series, NVIDIA halved the clock rate of the
shaders (they did this to help with power/heat issues I’ve
been told), so a lot more cores are needed to match a 500
series card for CUDA. Also, the 500 cards have much more memory
bandwidth available per core, so they’re more efficiently used.
The newer cards are faster for standard gaming 3D tasks, but not
CUDA, at least not for 32bit fp anyway.

Teddy, just a thought, should we include AE 12.1 results at all
in this thread? It might confuse people. Perhaps a new thread
for 12.1 data? Or at the very least ask posters to make it very
clear which version of AE they’re using.

Ian.

SGI Guru

[Ian Mapleson]

Joe Presswood writes:
> 45 seconds

I hate to sound sceptical, but that doesn’t sound right to me. Such
a result doesn’t remotely correlate with other metrics such as Arion:

https://www.randomcontrol.com/arionbench

Are you sure the render is not being cached or somesuch? ie. purge
the cache and all memory before starting the test, and of course
make sure it’s set to RayTraced3D, not Classic3D.

Can you post some screenshots of the NVIDIA Control Panel or
AE Preferences showing the CUDA GPU pool in use?

If it really is running that quick then it would be imperative
to find out why, but alas I doubt it’s correct. :/

Ian.

SGI Guru

[Ian Mapleson]

Interesting how the performance scaling occurs with this test,
or rather doesn’t (usage per GPU isn’t maxed), ie.:

4x GTX 580: 2 mins 41 seconds.

That is only just a little bit quicker than 3x 580.

Mbd is an ASUS P9X79 WS, 3930K @ 4.7, 64GB/2133 RAM, 128GB MAX IOPS
for the AE Cache.

OTOH, the result correlates quite well with Paul’s dual-690, given
the 680 isn’t as quick as a 580 for CUDA, so it makes sense that
4×580 would be a bit faster than 2×690.

Anyone here have 2+ Titans? That would be interesting.

I was wondering whether the higher PCIe bw of the newer P9X79E WS
would help here (x16/x16/x16/x16 for 4 cards, instead of x8/x8/x8/x8),
but who knows. Is there such a thing as a tool which allows one to
monitor the bandwidth usage of the PCIe links?

Also, have a look at the results for the Arion test, the systems
employed are kinda whacko (I’m thinking water-cooled 1-slot GPUs
for sure):

https://www.randomcontrol.com/arionbench

My tri-580 entry is at no. 26. I’ve tested 4×580, it gave a score
of 5317, which should take the no. 15 spot. 8)

Ian.

SGI Guru

[Ian Mapleson]

Just a thought, is it possible to configure the aep file so
that RayTraced3D mode is selected automatically when the aep
is loaded?

Ian.

SGI Guru

[Ian Mapleson]

> 3.6 wasn’t stable – moved the OC back to 3.4…all good now.

Pity you had to reduce it. Ah well.

> Turned off 4-way SLI and got 3 min..35 seconds.

Yup, SLI is not relevant for AE/CUDA.

Ian.

SGI Guru

[Ian Mapleson]

Paul Forcier writes:
>Hey Ian you nailed it. My CB 11.5 64 bit is:
>
> CPU – 17.13

😀

A linear extrapolation from my Dell/X5570 results suggested
17.05; I erred on the side of caution, but I forgot your
system has slightly faster RAM. Either way, an oc’d 3930K
will give about 14 or 15 at best, so 17+ is very nice.

> GPU – 37.27

Btw, nobody really bothers with CB GPU numbers anymore.
They stopped making any useful sense a long time ago.

Ian.

SGI Guru

[Ian Mapleson]

I guess it depends on what one is doing. Lots of cores in
one system will obviously be better for Classic3D mode,
or for rendering a single large image (I have a friend
who often does this, very high res for huge advertising
boards, roadside signs, etc.)

OTOH I too have commented on the idea of having 2 systems
so that while one is rendering one can still progress with
other work on the 2nd system.

As always it boils down to available budgets, intended
tasks, etc. I’m sure we’d all love a Titan and six Teslas
on a 7-slot mbd with uber water cooling. 😀

Ian.

SGI Guru

[Ian Mapleson]

Paul Forcier writes:
> Hi Ian, I’m using the Xeon e5645’s which were less expensive and easily
> overclocked, unlike the current generation of Xeon which are locked out
> in that regard.

Indeed! The pice looks quite good for a 6-core 2.4 X58. Wouldn’t make
sense for a single-socket system of course, but for a dual-socket setup
that’s a nice compromise.

> better chips of that generation like the 5690, though they were a great
> deal more expensive.

Yeah, exponentially explodes cost-wise above a certain point. That’s why
I initially bought a used X5570 4-core (by default it’ll run at 3.2)
which only cost me 200 UKP at the time. It was cheaper than an i7 950,
has 50% faster QPI and a higher TDP, so in theory should oc easier,
though at least on my board (Asrock X58 Extreme6) this has proven a
challenge with the board maxed out to 24GB. I did manage to get a crazy
cheapo i7 990X much later but I’ve not installed it yet, still testing
the X5570 setup.

> My system has been very stable with this OC and have had no issues. This
> effectively has given me 12 Cores at 3.4 Ghz easily and 24 threads with
> multi-threading.

That should run really well, akin to a top-end Dell T7500 with 2 of the
best 6-cores. What do you get for CB 11.5?

> core. They are 1600s. I’d have to check to see how fast I have them
> running with the OC.

I’ve read about issues trying to oc X58 boards when all the RAM slots
are populated. Be interesting to know what you ended up with.

> There are i7s out there with 6 cores that will provide a higher benchmark
> on the benchmark test, with a high OC well over 4 ghz,…

I’m not so sure about that. 😀 Yes, the 990X, 3930K, etc. can reach
4.7/5.0 respectively with a good setup, but I suspect your decent clocks
and 12 cores will mean your setup actually has a useful advantage. Not a
huge amount better, but definitely better. I’d be surprised if your CB
11.5 score was less than about 16 or 17. I’ll make a guess and say it’d
be about 16.5, 17 at best.

> I am considering adding another GTX 690 for SLI with some games for my
> kids (ok, maybe me too, such as Battlefield 3 which responds well with 4
> way SLI on the two 690s) …

You’ll need extra tissues to cope with all the drooling. 😀

> run the benchmark again and report in….

Should be interesting!!

> difference in my eyes led me to the older SR-2 which can still be found
> out there…

Yes, that was the conclusion I came to when talking to my friend. For an
oc’d dual-XEON setup, older X58-based boards make more sense.

Thanks for the info!!

Ian.

SGI Guru