Nvidia Announces Quadro M6000 Graphics Card With GM200 Maxwell GPU

[N.Broekhuijsen]

Just days after announcing the GTX Titan X, Nvidia is announcing a second graphics card that uses the GM200 Maxwell GPU. This one isn't geared towards consumers or prosumers, however.

Nvidia Announces Quadro M6000 Graphics Card With GM200 Maxwell GPU : Read more

 

[scook9]

Why cant we get more high-end consumer cards with rear (front?) facing power connectors? This would be a boon to small form factor boxes as well as slimmer HTPC builds.

 

[Blazer1985]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

 

[ohim]

The day when AMD/Nvidia will stop selling gaming GPUs with modified drivers at insane prices for professionals, it will be a good day !

Though Quadro/FirePro GPUs have some extra components, most of the card is just a gaming GPU with different drivers ..

 

[doron]

The day when AMD/Nvidia will stop selling gaming GPUs with modified drivers at insane prices for professionals, it will be a good day !

Though Quadro/FirePro GPUs have some extra components, most of the card is just a gaming GPU with different drivers ..

The majority of the money you pay for any chip, and almost any product, goes towards R&D and marketing.
Professional-grade gpus require extra man-power in drivers and support, and address a smaller market, which is probably the main reason for such price different.

 

[Colinmb123]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

Actually any program that is able to use Nvidia's HPC features can take advantage of it. We use the CUDA cores from a couple of Quardo K5000s for Ansys and Fluent to accelerate our solvers.

 

[Eximo]

Support being the key there. If I am a professional at a company with a pile of Quadro and Tesla cards, I can call up Nvidia and have them fix an issue I am having, or work directly with other software vendors to fix their code. Can't get that kind of support for gaming.

 

[Blazer1985]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

Actually any program that is able to use Nvidia's HPC features can take advantage of it. We use the CUDA cores from a couple of Quardo K5000s for Ansys and Fluent to accelerate our solvers.

You surely can but if your solver uses FP64 operations (and many, many do) you will be getting only 0.2Tflops against 1.3Tflops of current generation Quadro K5000 or Titan Black.
Looks like a deal breaker to me

 

[Colinmb123]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

Actually any program that is able to use Nvidia's HPC features can take advantage of it. We use the CUDA cores from a couple of Quardo K5000s for Ansys and Fluent to accelerate our solvers.

You surely can but if your solver uses FP64 operations (and many, many do) you will be getting only 0.2Tflops against 1.3Tflops of current generation Quadro K5000 or Titan Black.
Looks like a deal breaker to me

Wow. Just noticed that huge drop. Misread it originally. What they heck happened?

 

[Colinmb123]

Wow. Just noticed that huge drop. Misread it originally. What they heck happened?

Never mind. Found out that Nvidia doesn't want to waste the space to implement a 1/3 or 1/2 fp32 capable fp64 section. UGH. That sucks.

 

[dragonsqrrl]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

It will limit the market for this particular card, but you know there are plenty of "compute intensive" applications for fp32 performance in science, rendering, etc. Rendering in particular, those that rely on visualization such as in the film, game, or animation industries will be unaffected by the lack of double precision performance.

 

[ohim]

The day when AMD/Nvidia will stop selling gaming GPUs with modified drivers at insane prices for professionals, it will be a good day !

Though Quadro/FirePro GPUs have some extra components, most of the card is just a gaming GPU with different drivers ..

The majority of the money you pay for any chip, and almost any product, goes towards R&D and marketing.
Professional-grade gpus require extra man-power in drivers and support, and address a smaller market, which is probably the main reason for such price different.

I think you are wrong, since there have been mods in the past to gaming GPUs into these professional GPUs. If they put these teams to make an all around product that you can switch in the driver between gaming and CAD, you have a win win situation , more cards sold even at lower prices still bring more money than few cards at very high prices ... but this is just me.

 

[op8]

TIL, that Tom's had a sister site.

 

[Blazer1985]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

It will limit the market for this particular card, but you know there are plenty of "compute intensive" applications for fp32 performance in science, rendering, etc. Rendering in particular, those that rely on visualization such as in the film, game, or animation industries will be unaffected by the lack of double precision performance.

Yeah, that's what I thought too. But since I use unbiased renderers (which should use mostly FP32 instructions by their own admission) I took a look at many benchmarks and the unlocked FP64 titan was far ahead the gtx780 / 780ti. The old 580 and quadro equivalents were behind the titan but topped the 680s and 780s (fermi had more FP64 units than GK104 and on "big"kepler they could only be enabled by drivers). So.. looks that FP64 are surely important in rendering, science idk but if I were to write a solver I'd stick to the higher precision double float

 

[dragonsqrrl]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

It will limit the market for this particular card, but you know there are plenty of "compute intensive" applications for fp32 performance in science, rendering, etc. Rendering in particular, those that rely on visualization such as in the film, game, or animation industries will be unaffected by the lack of double precision performance.

Yeah, that's what I thought too. But since I use unbiased renderers (which should use mostly FP32 instructions by their own admission) I took a look at many benchmarks and the unlocked FP64 titan was far ahead the gtx780 / 780ti. The old 580 and quadro equivalents were behind the titan but topped the 680s and 780s (fermi had more FP64 units than GK104 and on "big"kepler they could only be enabled by drivers). So.. looks that FP64 are surely important in rendering, science idk but if I were to write a solver I'd stick to the higher precision double float

Really, what unbiased GPU accelerated renderer are you using? And with what 3D package? That's interesting because in my (admittedly limited) experience with GPU accelerated renderers, none of them use double-precision. I've tried FurryBall with Maya, and my roommate is starting to get into V-Ray RT with Maya, which as far as we can tell from test scenes and online benchmarks, also doesn't rely on double precision.

Just based on the evidence you gave in your comment, it's not necessarily indicative of the renderer utilizing an fp64 performance advantage. I know of several examples where the 580 outperformed the 680 in fp32 opencl and cuda workloads. The 580 outperforming the 780 though is a little strange, since it tends to outperform the 580 in both single and double precision workloads. It's possible you may be assuming causation where there is only a correlation in higher fp64 performance.

 

[Blazer1985]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

It will limit the market for this particular card, but you know there are plenty of "compute intensive" applications for fp32 performance in science, rendering, etc. Rendering in particular, those that rely on visualization such as in the film, game, or animation industries will be unaffected by the lack of double precision performance.

Yeah, that's what I thought too. But since I use unbiased renderers (which should use mostly FP32 instructions by their own admission) I took a look at many benchmarks and the unlocked FP64 titan was far ahead the gtx780 / 780ti. The old 580 and quadro equivalents were behind the titan but topped the 680s and 780s (fermi had more FP64 units than GK104 and on "big"kepler they could only be enabled by drivers). So.. looks that FP64 are surely important in rendering, science idk but if I were to write a solver I'd stick to the higher precision double float

Really, what unbiased GPU accelerated renderer are you using? And with what 3D package? That's interesting because in my (admittedly limited) experience with GPU accelerated renderers, none of them use double-precision. I've tried FurryBall with Maya, and my roommate is starting to get into V-Ray RT with Maya, which as far as we can tell from test scenes and online benchmarks, also doesn't rely on double precision.

Just based on the evidence you gave in your comment, it's not necessarily indicative of the renderer utilizing an fp64 performance advantage. I know of several examples where the 580 outperformed the 680 in fp32 opencl and cuda workloads. The 580 outperforming the 780 though is a little strange, since it tends to outperform the 580 in both single and double precision workloads. It's possible you may be assuming causation where there is only a correlation in higher fp64 performance.

I use Octane renderer but cycles and iray perform very similarly. Here's an example from tom's itself
IIRC gtx780 = titan and gtx780ti = titan black, the only difference being the possibility to unlock FP64 from the panel.
http://media.bestofmicro.com/Y/8/386144/original/19-CUDA-3ds-Max-iRay.png
Although 680 and 780 pretty much crushed 580 in FP32 with gpgpu the old fermi got his revenge :-D
What do you think?

 

[dragonsqrrl]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

It will limit the market for this particular card, but you know there are plenty of "compute intensive" applications for fp32 performance in science, rendering, etc. Rendering in particular, those that rely on visualization such as in the film, game, or animation industries will be unaffected by the lack of double precision performance.

Yeah, that's what I thought too. But since I use unbiased renderers (which should use mostly FP32 instructions by their own admission) I took a look at many benchmarks and the unlocked FP64 titan was far ahead the gtx780 / 780ti. The old 580 and quadro equivalents were behind the titan but topped the 680s and 780s (fermi had more FP64 units than GK104 and on "big"kepler they could only be enabled by drivers). So.. looks that FP64 are surely important in rendering, science idk but if I were to write a solver I'd stick to the higher precision double float

Really, what unbiased GPU accelerated renderer are you using? And with what 3D package? That's interesting because in my (admittedly limited) experience with GPU accelerated renderers, none of them use double-precision. I've tried FurryBall with Maya, and my roommate is starting to get into V-Ray RT with Maya, which as far as we can tell from test scenes and online benchmarks, also doesn't rely on double precision.

Just based on the evidence you gave in your comment, it's not necessarily indicative of the renderer utilizing an fp64 performance advantage. I know of several examples where the 580 outperformed the 680 in fp32 opencl and cuda workloads. The 580 outperforming the 780 though is a little strange, since it tends to outperform the 580 in both single and double precision workloads. It's possible you may be assuming causation where there is only a correlation in higher fp64 performance.

I use Octane renderer but cycles and iray perform very similarly. Here's an example from tom's itself
IIRC gtx780 = titan and gtx780ti = titan black, the only difference being the possibility to unlock FP64 from the panel.
http://media.bestofmicro.com/Y/8/386144/original/19-CUDA-3ds-Max-iRay.png
Although 680 and 780 pretty much crushed 580 in FP32 with gpgpu the old fermi got his revenge :-D
What do you think?

I don't know. It's strange, but to me it doesn't seem like the Titan has enough of a performance advantage to signify that the renderer is relying on double precision. In fact the performance gradient of the Kepler cards just looks like what you would expect from a typical fp32 workload. The only strange thing is the 580 which shouldn't have remotely close to the Titan's theoretical single or double precision performance. But even that's not too unusual when you look at the real world performance results from other renderers, I just can't make heads or tales of them sometimes. I've found that it's quite common in these types of workloads to find strange performance outliers for no apparent reason. For example the Titan outperforms the K6000 in FurryBall. It may just come down to quirkiness involving optimizations, or a renderer favoring a certain architecture for some reason. I could be wrong, but I doubt that the Titan and 580 are leading in Iray due to a reliance on double precision. Add to that the fact that Nvidia is advertising Iray for the M6000 and their just announced Maxwell based Quadro VCA, which just wouldn't make sense if the renderer relied on double precision.

 

[Blazer1985]

Again I agree with your theory but I find hard to justify the jump you get by only switching on the FP64 on the titan. Mmmmh I'll search for some clues in the octane forum.

 

[mctylr]

It will limit the market for this particular card, but you know there are plenty of "compute intensive" applications for fp32 performance in science, rendering, etc.

I can't comment on renderers, but what scientific CUDA or GPGPU applications use fp32? Honestly, I don't know of any applications where fp32 would be suitable.

 

[Colinmb123]

I can't comment on the rendering concerns between fp32 vs fp64 and the performance differences in analytic instances is not the only driving factor. Precision and range are what are the desired advantages for using fp64 in my applications. Killing the performance of fp64 capabilities would definitely be a hindrance for our structural and CFD analysis.

 

[mapesdhs]

0.2Tflops FP64? Is Nvidia kidding with this quadro?
That's an awesome card for viewport performance but of no use in compute intensive tasks.

Then buy a Tesla. That's what NVIDIA wants such customers to do, and for the
intended target market, it makes sense, given the features that Tesla has which
the gamer cards don't. Having said that, I do think the FP64 is erring a tad too
much on the weak side for this round of updates. Teslas are good, but expensive.
At least with the original Titan there was a middleground suitable for developers,
if not for end product deployment.


Re renderers, anyone used Arion? And AE is a perfect match for FP32-only compute.

Also, gamer cards lack ECC, full speed return path and caching features of Quado/Tesla
models. This makes all the difference for some tasks. Note the 580 is strong for CUDA
for various reasons: 2X faster shader clock, big mem bw per core, etc. I talked a fair bit
with C. Angelini about this, as it seemed weird at first that the 580 remained so strong
for CUDA, but the details of how it works makes it less surprising. It beats all of the 600
series for most CUDA tasks, and the only 700 series card which can beat it consistently
(though not always) is the 780 Ti (and Titan), but the cost advantage pf used 580s
means the 3GB continues to be a strong solution for budget-starved solo pros looking
to build a good machine for AE. My test unit with four 580 3GB is faster than two Titan
Blacks (Arion bench, AE, etc.)

Ian.

 

[Mac266]

That cooler though...