buying a tesla card

[leo.surrealismo]

Has anyone used a tesla card? Does it worth buying one to use it with vray 2.0? or would it be better buying 2 480gtx?

thanks in advance,

leo

[dpmitchell]

Tesla cards are incredibly expensive. If you are using the card for viewport performance, the gaming cards (GeForce) do not compare to the quadro cards (With the new Quadro 4000 card available for ~ $700). As a matter of fact, as GeForce cards go, the viewport performance reports seem to indicate that the GeForce GTX 400 series cards are a major step backwards in viewport speed from the GeForce 8-9 series cards.

If you're using the card for rendering, like with Vray RT GPU, then maxing out the processor cores seems to make the most sense. Bang for the buck. GPU rendering performance = # of cores x core clock speed.

Tesla 2050 = ~$2500; 448 CUDA cores.

GTX 470 = ~$250; 448 CUDA cores.

So for $1200, you could snag a Quadro 4000 and two GTX 470 cards, and you'd have a viewport killer Quadro, along with 896 CUDA cores to dedicate strictly to rendering. Hell, you might find that dropping down to the GTX 460 is more cost effective, as long as you have enough PCI slots. They should NOT be in SLI mode anyway for rendering.

One real difference between the Tesla and GeForce cards, according to nVidia, is that nVidia oversees the production of the Tesla cards, warranties them in server conditions for 3 years, and says they are designed for the 24 hour stress of GPU computing, whereas the GeForce cards are made by third parties, and may fail under heavy stress conditions resulting in work loss. Well for 175-250 bucks, I'll ride those GeForce cores as long as they'll go before I stick a $2,500 tesla card in here for reliability that will be obsolete well before the warranty expires. Also, keep in mind that certain models of tesla cards, like the tesla M, are not warrantied for workstation use, only server use.

Also keep in mind, Vray RT GPU is an active shade renderer, not a production renderer. I believe the production renderer of Vray 2.0 is still CPU based. If you have the subscription Advantage pack for Max, it includes the iRay renderer, which IS a GPU based production renderer...an idea of where Vray will go in the future.

[leo.surrealismo]

thank you for the clear explanation.

If I run out of pci express slots, is there a way to connect more video cards externally thru an expansion slot or something else?

[Donald2B]

No, but you can distribute render with VRayRT. So if you have another computer, load it up with some video cards and you are good to go!

[dpmitchell]

More analysis on Tesla vs. GeForce GTX 400 series for rendering.

Nvidia cut a lot of features from the GTX 400 version of the Fermi architecture, but left FP64 (double precision floating point, 64 bit) on the GTX 400 cards simply so developers could run testing to improve the Tesla line. But they limited it to 1/3 of the blocks of CUDA cores, so FP64 on the GTX 400 series is effectively limited to 1/8 of the single precision floating point operations per second (somewhere around 900-1000 single precision GFLOPS for a GTX 470). The Tesla/Quadro FP64 is much higher than that, possibly closer to one half of the single precision numbers.

Even so, at 1/10 the price of the Tesla 2050 card for a GTX 470, and with single precision GFLOPS nearly identical to the Tesla 2050, not enough of a disparity to warrant the Tesla on that alone. For comparison's sake, the GTX 470 absolutely smokes most quad cores in GFLOPS (with massively overclocked i7 930's...like @ 4.5 ghz...getting somewhere in the 65 GFLOPS range) anyway.

Then, if you look at CUDA cores and clock speed:

Tesla 2050 with 448 CUDA cores runs at 1.15 Ghz; 3 GB 384-bit GDDR5

GTX 470 with 448 CUDA cores runs at 1.35 Ghz; 1.28 GB 320-bit GDDR5

Core speed total goes to the GTX 470. I am not sure how much of an effect the dedicated memory has in terms of rendering, because the whole concept of GPU rendering (particularly global illumination) is so new. Maybe someone has some input on this.

So really, for our purposes, this is what Tesla all comes down to (from Nvidia):

Zero error tolerance stress testing
Stress tested in manufacturing for several days
Manufactured by NVIDIA and guaranteed to be of highest quality
Added margin in memory and core clocks for added reliability
Enterprise level support
Three year warranty, higher priority bug resolutions and feature requests, and product lifecycle of 18-24 months

http://www.nvidia.com/object/why-choose-tesla.html

Tesla also has ECC enabled (pre-emptive error detection), which has been disabled on the GeForce GTX 400 cards, as they don't need it for gaming purposes, but it is more important for GPU computing. But for RT rendering, I am just not that concerned with that for the additional money. Nvidia is banking on the fact that people will shy away from the GTX 400 fearing it would ultimately burn out or at least encounter momentary failures when placed under the stress that GPU computing would place on it. They advocate Running the 3ds Max viewport from a Quadro, and running a Tesla for GPU computing (or rendering) in iRay.

Based on the specs, two GTX 470 should...should....far outperform the Tesla 2050 in terms of benchmarking for RENDERING only. The longevity is a whole separate issue, but I can't imagine these cards will burn out under reasonable conditions. If they run at 100% capacity 24/7 in big production houses, yes, but otherwise, I bet they do just fine.

Also note, at the bottom of Nvidia's Tesla page, the following asterisk:

Most benefits listed above are also available on Quadro product family.

Exactly...as Fermi goes, Tesla is a basically overpriced Quadro, which has, in the past, been an overpriced Geforce with special drivers and without the limitations Nvidia puts on the Geforce. Although I would never think of running the viewport with anything other than a Quadro. Too frustrating, slow, and just asking for some poor inanimate object to have an unfortunate "accident" I have a Quadro 4000 (fermi) on the way, so hopefully it runs as smooth as they've been hyping that it does!

[dpmitchell]

BTW, I will be adding a GTX 470 to this new build for purposes of testing the consumer fermi in GPU rendering conditions, and comparing it's performance to the Quadro 4000.

I will run the ever loving hell out of that GTX 470 just to see how far I can push it and see how it performs, so I will report once I get it complete. The fun begins this Saturday!

The build will be:

ASUS Rampage III Formula with Sata III (6 gbps)
Crucial C300 SSD Sata III (6 gbps)
12 GB Triple Channel DDR3 1300 RAM
Nvidia Quadro 4000 (Fermi)
Nvidia GTX 470 (Fermi)
Intel i7 930

[leo.surrealismo]

I'll be expecting the results! thank you

[Donald2B]

I wonder how the newly announce GTX580 will do?

[dpmitchell]

I wonder how the newly announce GTX580 will do?

Great, I'm sure. For almost the same price, it seems to be far more powerful than the GTX 480. But as I've been discussing with Valdo in the other thread, I just can't justify the price jump from the GTX 470 for our uses.

If I'm running SLI and using it for Direct X (gaming), or if I'm using it for rendering and I have no possibility of using rendernodes so I'm maxed at the PCI slots on just one motherboard, then it would make sense.

For us, however, the GTX 470 hits the sweet spot for price and performance. (GTX 460 is even tempting because of the processor clock speed and sub $200 price point, especially with sales, but for just a few $ more, the 470 has 14 multi processors compared to 7 on the 460). Gotta love distributed rendering I bet gamers wish they could do that

[leo.surrealismo]

How are you doing with your new quadro4000? Did it fulfill your expectations?