hello evryone,
i want to know if “cycles render” is conpatible with Nvdia gtx590 on sli? or with a Nvidia tesla card (like tesla c1060…)…
:o anyone have an experience with this config?
Not sure on Tesla… SLI defintely isnt made use of at the moment… afaik there isnt even a way to pick a secondary card, which is a shame as I was concidering using one gpu for gaming and one for rendering… as at the moment, its either/or.
afaik there isnt even a way to pick a secondary card,
From todays developers meeting, multi gpu rendering should be ready by the end of the month
Not sure on Tesla… SLI defintely isnt made use of at the moment… afaik there isnt even a way to pick a secondary card
=\ it’s a pity… the render of cycle good but I think it is too slow… for render a complex scene or fast animated a project…
HAAAAAaa
From todays developers meeting, multi gpu rendering should be ready by the end of the month
it is a good news!
it is sure?
Awesome… finally I can play games while I render away, great news, thanks.
xD it’s the ideal of the gpu technology!
else nothing for tesla?
blender with tesla? :spin: no? 
because a mate will sold me a tesla card c1060 … at a very good price… = (( but if it s not compatible with blender … i will use max for a render :\ …
Blender should work with a Tesla, a GeForce and a Quadro as long as the CUDA driver is installed.
As soon as there’s a Linux 64 bit multi GPU build I´ll be able to test Cycles on 4*Tesla 2070.
However, unless Cycles uses double precision, which I doubt, a GeForce should be faster, as Teslas have lower clocks in the shader units. The huge advantage of Teslas are the ECC Memory and the amount of memory for precise scientific applications.
And to clearify it for the 100th time, SLI is not multi GPU.
No luck, the Teslas sit in a datacenter, have no screens and no GUI installed. They’d hate me if I start to install X.org and start refurnishing with screens and stuff 
And to clearify it for the 100th time, SLI is not multi GPU.
but SLI it is a cumulat of gpu card… it isn’t same?
No luck, the Teslas sit in a datacenter, have no screens and no GUI installed. They’d hate me if I start to install X.org and start refurnishing with screens and stuff
Damned … i read that the tesla must be use with an api cuda … =\ it’s same for windows seven… tesla is a story for C developper
I know SLI is the Nvidia way to make two or more GPUs work as a single device, so theoretically your system should only see one logical GPU. I guess that’s the reason you say that SLI is not multi GPU.
But GTX590 is actually two GTX580s embedded in a single board (kind of SLI implementation on a single PCX port), and it has already been stated somewhere in these boards that Cycles can only use one of the two GPUs in a GTX590 (not detecting both as a single device), so I suppose it’s safe to assume that SLI setups will face the same problems a multi GPU rig without SLI would face when working with Cycles. Am I wrong?
SLI:
Only works with GPUs of the same type. It also requires a mainboard with a chipset supporting SLI as the data between the two cards is sent over the bus. The SLI bridge is something like a bandwith extender as the PCIe bus alone would be too slow for the data transfer.
It works with 2, 3 or 4 cards.
Multi GPU in Raytracing:
It works with different GPUs and always takes the memory of the smallest card. The scene data should be transmitted to each of the cards so all the data is on the card and you only have operations on the card and never between card and system memory as it would work over the system bus which is a dead slug compared to the on-card memory bus.
Each of the GPUs now can access it´s assigned portion of the scene and render it.
It works with 2-8 cards, even more if you got e.g. an extender box for multiple cards you attach to your PCIe.
–
As far as I know you can use e.g. a 590 for multi GPU rendering but not without efford, as I said, SLI cards and dual GPU cards are useless abominations, I don´t follow them very closely.
–
Lastly, the c1060 is going to suck.
It has 240 streaming processors at 1296 MHz resulting in:
SinglePrecision: 936 GFLOPS
DoublePrecision: 78 GFLOPS
However it has 4 GB VRam but costs ~1800 Euro.
A GTX580 has 512 streaming processors running with 1544 MHz.
This results in an FMA (floatingpoint multiply add) SP performance of 1580 GFLOPS.
The Double Precision performance is capped at 1/8th of the Tesla, but I think Cycles runs in SP.
However it only has 1.5GB memory per default, but there are a lot of versions with 3GB already and you get it for about 500 Euro.
A tesla S2050 (4 cards) would have brutal 4120 GFLOPS in SP… it costs whoopdidoo: 11.000 Euro.
I was vaguely aware of most of what you said about the differences between SLI and multi GPU rendering. I suppose multi GPU support for Cycles will be very similar to the one in Octane.
I was planning myself to get a GTX590 hoping I could use both of its GPUs when multi GPU support gets implemented in Cycles. Its performance will be probably less than two GTX580s, but it would be cheaper and will consume less power. But then I took an arrow to the knee.
Well, not really (sorry, I couldn’t resist). I just learned that those 3 GB of onboard memory that the GTX590 sports are actually split between the two GPUs, so you only get 1.5 GB available for your scene.
Since I’ve already reached the limit in some of my scenes where my Quadro 4000’s 2 GB aren’t enough, 1.5 GB is definitely a no go.
I suppose I will have to start checking those GTX580s with 3GB you mentioned.
A GTX590 should be faster than 2 GTX580 because of reasons! (9gag?)
It seems like a GTX590 works flawlessly with CUDA, no guarantee for Cycles and that you can use the 3GB:
Firstly the two GPUs seem to be able to transfer between their memory onboard, which bypasses the system bus a few GB/s compared to GPU/VRAM communication a few hundred GB/s.
Secondly it seems since CUDA4.0 there is unified virtual addressing. In CUDA each card has it’s own memory, it’s not one superdevice, CUDA sees it as 2 independent cards and can address the whole 3 GB on it.
So in theory 2*GTX590 should give you ~5000 GFLOPS SP with 6GB of memory, if cycles is coded to support this unified virtual addressing.
Cost would be 1200-1500 Euro for those 2 cards and prolly another 200 for watercooling, else there’s not enough room.
Not exactly a price where you want to experiment but surely better
More on that:
Or wait 2 months for the GF700 and see what happens then 
Cost would be 1200-1500 Euro for those 2 cards and prolly another 200 for watercooling, else there’s not enough room.
I think we should also include a power supply 900 to 1100w ~ 300€… or the system is going to insult (just crash!)
=> ~ 1500 to 1800€ … i take a bullet in my leg
this would not be better to buy 4 x gtx580 with 3000mo? :spin:
wait, I will do a calculation price / quality / consomation
I’m not sure if I got this right. If Cycles gets coded to support the unified virtual addressing in CUDA 4.0 (and that’s a big IF), will I be able to fit a 3GB scene into a GTX590? Or will it work as two GTX580s with 1.5 GB each, thus limiting the complexity of the scenes you can render? Someone in the thread you linked asked a similar question, but the answer dind’t seem very clear to me.
If this article at Tom’s Hardware is to be believed, I think we will have to wait longer than that. According to this, performance versions won’t be available until Q3. If you are kind of a prudent person, like me, you’ll also wait a bit longer for the first hands-on reviews to appear before even considering buying new hardware. So, I’m assuming it would take at least 9 months for me to get a top of the line Kepler based card. I think I’d rather settle now for a single GTX580 (or GTX590 if I can fit 3GB scenes on it) for the time being and make do with it. In a year from now I would be more than ready to get an additional Kepler card if it proves to be worthy.
When I built my rig I had this kind of future upgrade in mind and chose a Corsair CMPSU-1200AXEU (1200W) for my PSU, so I think I have that area kind of covered.
helo!
When I built my rig I had this kind of future upgrade in mind and chose a Corsair CMPSU-1200AXEU (1200W) for my PSU, so I think I have that area kind of covered
good idea!
…
my calculation suck for GTX 580… it’s maybe an error… :spin:
But anyway, there are and other alternative but to buy gold price of the GPU … i look for renderfarming online … “green button” have a addon for blenderand can be render a complexe final render by gpu… but we should cost the processor by hour…
check that => http://www.greenbutton.net/Applications/Blender
“Generally we boost the initial number of cores by 400% but please note, the extra cores can take a further 15 to 20 minutes to come online, so small jobs (less than 100 frames or single frame renders) may complete before the additional cores are available.”
…
" For a limited time we are offering a special of just US $0.18 per core hour for all Blender, Yafaray and LuxRender jobs. This special is available until the 31st of Janurary 2012. "
What do you think plz? 
Easy math, as special feature they offer 1 CPU CORE! hour for 0.14 Euro
A GTX580 with an optimized raytracer should performs around 400 times faster than a singlecore.
So what a GTX580 does in one hour costs you 56 Euro minus the 0.375kWh electricity.
Here it’s .15 Euro/kWh.
A GTX580 3G costs you 500 Euro.
500 / (56 - 0,15*0.375) = 8.93h
If you have to render over 9h on one core at the renderfarm it’s cheaper to render at home wiht a GTX580. If you got the time.
Obviously, as the GTX580 is 400 times faster than 1 CPU-core, 800 Cores are two times faster than 1 GPU. However you could split the rendered frames into Frames 1-100 and 101-200 and render on 2 GPUs simultaniously.
Let’s take an 10 minutes animation 1080p25
256010 = 15000 frames.
A quick render with a quadcore at home averages at 15 minutes per frame.
So a single core would take 1h per frame.
Totals in 15000h or 625days, or almost 1.71 years 
We need it in one hour.
So let’s rent 15.000 cores, which is most likely not possible it’ll take some hours for sure but for the ease of the math.
15.000 cores would take 1h to render the 10 minute animation.
15.000*0.18 Euro = 2700 Euro to render the Animation.
You could buy 4 GTX580 for that.
1 CPU core takes 1h, 1 GTX580 should be 400 times faster, so it renders 1 frame in 9 seconds.
15.000 frames split on 4 cards are 3750 frames per card, 9 seconds per frame is a total of 9.38h for the animation.
Those guesstimations strongly depend on the speed advantage of the GPU over the CPU though.
400x is a pretty much ideal figure. It’s most likely to be less.
But you can just take one of your frames and compare for instance in cycles with CUDA and with CPU on 1 core and measure how much faster it is, then do the math.
I am positive that if you have a smaller studio and lots of animations to render you’ll reach the point where a GPU is more economical pretty soon.