Nvidia Announces Quadro K6000 Workstation GPU
Today at SIGGRAPH 2013 in Anaheim, CA, Nvidia announced the new 'top card' for its Quadro line of workstation graphics cards, the Quadro K6000.
Nvidia is today using SIGGRAPH 2013 as a launchpad for a new 'top card' in its Quadro line of workstation graphics cards. As the new flagship of Nvidia's professional graphics line, the Quadro K6000 introduces unrivaled power for a workstation GPU as well as 12 GB of memory, the largest amount of GPU memory yet available. Check the specs below:
| SPECIFICATIONS | |
|---|---|
| GPU | GK110 |
| GPU Memory | 12 GB DDR5 |
| Memory Interface | 384-bit |
| Memory Bandwidth | 288 GB/s |
| CUDA Cores | 2880 |
| Single-precision Floating Point Performance | 5.2 TFLOPs |
| System Interface | PCI Express 3.0 x16 |
| Max Power Consumption | 225W |
| Thermal Solution | Ultra-quiet active fansink |
| Form Factor | 4.4”H × 10.5”L, Dual Slot, Full Height |
| Display Connectors | DVI-I DL + DVI-D DL + 2x DP1.2 + Stereo |
| Max Simultaneous Displays | 4 |
| Max DP 1.2 Resolution | 3840 × 2160 at 60 Hz |
| Max DVI DL Resolution | 2560 × 1600 at 60 Hz |
| Max DVI SL Resolution | 1920 × 1200 at 60 Hz |
| Max VGA Resolution | 2048 × 1536 at 85 Hz |
| Graphics APIs | Shader Model 5.0, OpenGL 4.3, DirectX 11 |
| Compute APIs | CUDA, DirectCompute, OpenCL |
Pretty impressive specsheet and the 2880 CUDA cores and 12 GB of GPU memory ensure it's not just the 'Quadro version of the Titan.' Nvidia is going to great lengths to present the advantage of the 12 GB of memory on the card in three key markets: animation and visual effects, automotive and product design, and energy exploration.
First, the company gave a prototype card to the boys (and girls) at Pixar. You can see what they had to say about it above. Considering Pixar doesn't normally endorse any kind of products, this is high praise. The particular interest in this card for the entertainment market is that they can use higher levels of detail than they previously could and still maintain interactivity.
Then Nvidia gave one to the design center at Nissan North America. The Pathfinder above is forty million polygons. Normally, when the design center receives a car design from engineering to do visualizations, they get a NURBS-based CAD model which they first convert into polygons and then spend hours (more than likely days) reducing the resolution of in order for the model to be useable. Things like reducing the overall polygon count without losing the fine detail, greatly reducing the complexity of the interior when doing visualizations of the exterior (and vice versa), greatly simplifying the chassis parts, etc. With the Quadro K6000, they were able to take the above model , import it and do elementary cleanup, and then surface it for rendering. This is instead of spending large amounts of time just making the model lightweight enough to manipulate in RTT Deltagen, Nissan's application of choice.
Lastly, one was given to Apache Corporation. Apache uses Terraspark's InsightEarth for oil and gas exploration. This work involves huge datasets - hundreds of gigabytes - and they are using GPUs to process the data. The larger GPU memory size means they can load much larger segments of their dataset onto the GPU, thiu greatly speeding up their work.
Nvidia expects the Quadro K6000 to be available this fall throughout its normal professional graphics distribution channels. Expected price is unavailable at this time, but Nvidia expects top announce its price as it gets closer to its ship date.




Preeeeety far away from any mere mortal and us "herculean small heroes" of gaming.
*Herculean small heroes? Yes. We are the toughest around in the playground, but the big guys are in the principal's office.
A TDP of 225W is mighty impressive when you consider that the Titan is rated at 250W.
However Gpgpu often perform better when not connected through sli cable :-)
A TDP of 225W is mighty impressive when you consider that the Titan is rated at 250W.
Well pro cards like quadro often clocked much lower than their geforce counterpart for stability so it's not really suprising if quadro cards like K6000 consume much less.
So the yeild finally good enough for nvidia to start selling product based on fullly enabled GK110.
However Gpgpu often perform better when not connected through sli cable :-)
You're essentially paying for the drivers, and the additional optimization and validation for a wide range of professional applications. Unless the "3D studios" you're referring to run games, I can't imagine why they would be running any sort of SLI setup in their workstations. None of the content creation or video editing software I'm familiar with can take advantage of SLI for viewport or output rendering acceleration.
And I don't think SLI plays much of a role in multi-GPU compute performance. I always thought SLI was meant specifically for rendering, pretty much exclusively in games. But Tesla cards also include SLI bridge support, so I'm not sure. As far as I know you don't need to be running two GPU's in SLI to achieve parallel scaling from the second GPU.
A TDP of 225W is mighty impressive when you consider that the Titan is rated at 250W.
Well pro cards like quadro often clocked much lower than their geforce counterpart for stability so it's not really suprising if quadro cards like K6000 consume much less.
The thing is the stock clock on the K6000 is estimated to be higher than any other GK110 based card, including the GTX780. In addition it uses a fully enabled 15 SMX GK110. It would be interesting to compare real world load power consumption between the Titan and K6000. If I had to guess the official TDP isn't telling the whole story on power consumption. One theory might be that the higher TDP on the Geforce cards allows for higher and more consistent boost-clocks and greater overclocking headroom, where as Quadro cards don't have a boost clock. It isn't unusual for a Titan to run consistently at or above 900MHz at load, despite the top boost clock of 876MHz in the specs.
In any case this card is going to be an absolute beast. 5.2 TFLOPS SP is very impressive, but 1.7 TFLOPS DP is just insane. It's been a long time since Nvidia launched a Quadro card with raw clocks and specs similar to its Geforce counterparts. They must be really confident in these top binned GK110's.
NVIDIA will sell far more of these - my wild guess is $4,650- than we mere mortals might expect, as the visual aids for the advancing realm of the Personal Supercomputer.
The recent increase in availability of of 4X and 8X CPU, 8 and 10 core CPU's (10-core Xeons are $3,500+) supporting up to 4TB of ECC RAM - that's a $35,000 story in itself- plus the important advances offered by PCIe coprocessing modules such as Tesla K20 and Intel Xeon Phi will bring> atmospheric, oceanic, astrohysical, chemical, metallurgical, and particle modeling, CGI feature film animation/ editing, gas, dynamic structural, thermal, mechanical, financial, resources exploration, and aerodynamic simulation to a new level of accessibility- well relatively.
The Titan supercomputer at Oak Ridge, currently the most powerful calculating machine in the World, is using massively parallel components of which the bulk of the fundamental components > AMD CPU's and the Tesla K20X coprocessors are a $97,000,000 configuration of multiples (18,600+ of each) of components found at newegg and amazon. That's the key to all this- string together many, semi-consumer level parts into a massive system > it's AI Modular Man! And the Quadro K6000 as an extreme performer will be one of the big picture windows to those events.
BambiBoom >
Quadro's > FX 4800, FX 580, FX 570, FX 550 (retired) > Aspirational > K4000
PS > And yes, the K6000 is guaranteed to be SLI- capable as are the Quadro 6000 and K5000 along with the Quadro FX 4800 (1.5 GB) and FX 5800 (4GB) of two generations ago.
There is NO sense in changing from nurbs purely to do rendering though, if that's what you've been doing (perhaps between formats) - especially when DeltaGen uses nurbs anyway.
Preeeeety far away from any mere mortal and us "herculean small heroes" of gaming.
*Herculean small heroes? Yes. We are the toughest around in the playground, but the big guys are in the principal's office.
sli will only work on games from what i know, DCC apps do not support sli.
Ya, that's not SLI.
In a Maximus configuration the Tesla handles parallel processing, freeing up the Quadro to devote all of its resources towards rendering. It works differently, and serves a completely different purpose.
Not that I know of.
to feed all those cores fast enough for the real CUDA potential to be fully exploited.
I just don't get why NVIDIA keeps using such narrow bus widths/speeds. Check
the CUDA page for the 760 review, the very high bandwidth per core is why the 580
is still so good for CUDA, despite having massively fewer cores. This thing really
needs a 768bit bus.
I don't know why so many keep asking about SLI. Of course the card supports
SLI, upper-end Quadros always have, but only certified mbds/chipsets allow this
to function, and even then OEM driver configurations can further limit what will
work, eg. I have a Dell T7500 and a couple of Quadro FX 5500s which should
run SLI, but they can't because Dell never supported it even though NVIDIA says
it should be ok (NVIDIA told me they had no idea FX 5500 SLI didn't work in a
T7500). In reality though, not many pro apps are written to properly exploit SLI.
Ian.
> I'd assume that the wider bus is massively expensive...
Not at all, it's been done before; the 280 had a 512bit bus, and NVIDIA already has the
expertise to do 1024bit bus tech.
> ... Are you saying that a 580 (570??) is competitive with a modern 760 in CUDA?
YES! Read the 760 review, it matches or beats a 780 much of the time. And two 580s
leave a Titan in the dust. See:
http://www.tomshardware.com/reviews/geforce-gtx-760-review-gk104,3542-19.html
NB: Fluidmark produces a composite score; only part of the test is CUDA, the rest is
normal 3D stuff which of course is much quicker with newer cards.
See this AE benchmark thread for example CUDA results from a wide variety of
GPU configs:
http://forums.creativecow.net/thread/2/1019120
My tri-580 run is currently the fastest submitted result.
Ian.