nVidia announces K6000 - almost fully enabled K110

[JeremyCT]

More CUDA cores than Titan and twice the buffer.

2,880 Cores, that's 15 of the 16 possible SMXes enabled. Titan uses 14. I don't suppose they'll release any fully enabled GK110 cards if they can't even do it for the professional market.

http://www.tomshardware.com/news/Nvidia-Quadro-Workstation-Specs-Release-Date-Price,23639.html

No release date or price yet. What do you suppose the odds of them getting enough of these chips to launch a consumer level card with that chip are? Volumes on Quadro cards are low, so easier to meet demand.

 

[EarthDog]

Thats a quadro... MEH! We won't see any iteration of this in a non Quadro form I do not imagine.

 

[Culbrelai]

GK110 is a derivative of the Tesla K20X, neither a Quadro (business graphics) nor a GeForce (Gaming graphics) but Teslas are used for GPGPU computing.

So this is essentially, a business version of the Titan.

Fascinating.

Can Quadros be SLI'ed?

 

[EarthDog]

GK110 is K20X. The lesser cards are derivatives if you are looking at it like that.

http://www.anandtech.com/show/6446/nvidia-launches-tesla-k20-k20x-gk110-arrives-at-last/3

 

[JeremyCT]

GK110 is a derivative of the Tesla K20X

That's kinda backwards. The GK110 was designed primarily with HPC and CUDA compute applications in mind, it has a lot more die area dedicated to features that gamers don't need or use, so it's the better chip to use where those features are valued and desired. K20x uses 14 of 16 SMXes, same as the Titan. This is the first K110 announced with this many cores enabled. I'm not sure what configuration the supercomputer customers got, but I'd guess it was probably the 14 SMX version.

Quadro customers had been "stuck" with GK104 cards until now. It's the 15 of 16 SMXes enabled bit that got my attention though. I know nVidia had more problems with yield at 28nm than the other major players, perhaps this is a sign that those issues have now (finally) been mostly sorted?

 

[Janus67]

GK110 is a derivative of the Tesla K20X, neither a Quadro (business graphics) nor a GeForce (Gaming graphics) but Teslas are used for GPGPU computing.

So this is essentially, a business version of the Titan.

Fascinating.

Can Quadros be SLI'ed?

I don't think you gain anything by using SLI for GPGPU.

 

[petteyg359]

Can Quadros be SLI'ed?

I don't think you gain anything by using SLI for GPGPU.

I would think the SLI "bus" would be an extreme bottleneck when dealing with the massive concurrency these are used for. A Radeon 7950 could theoretically hit 240 GB/s, and nVidia lists 288.8 GB/s for Titan. Even in PCIe 4.0, you only get close to 2 GB/s per lane.

 

[JeremyCT]

The Tesla K10 uses 2x GK104 chips. There must be some sort of concurrency for GPGPU applications there, but I'm not entirely sure about how things work in that sort of application.

 

[EarthDog]

I would think the SLI "bus" would be an extreme bottleneck when dealing with the massive concurrency these are used for. A Radeon 7950 could theoretically hit 240 GB/s, and nVidia lists 288.8 GB/s for Titan. Even in PCIe 4.0, you only get close to 2 GB/s per lane.

THe PCIe bandwidth and ram bandwidth are different, right? In that the ram bandwidth is internal on the card versus the data that needs to travel through the PCIe lanes... I'm not sure.

 

[Culbrelai]

The Tesla K10 uses 2x GK104 chips. There must be some sort of concurrency for GPGPU applications there, but I'm not entirely sure about how things work in that sort of application.

So it's like a GPGPU verison of a GTX 690? How do the two GPUs communicate? Do they even need to? Teslas don't actually push any pixels per se (for they don't have any video outputs) so they would likely not be able to SLI... but Quadros...

Either way really intruiging stuff.

That's kinda backwards. The GK110 was designed primarily with HPC and CUDA compute applications in mind, it has a lot more die area dedicated to features that gamers don't need or use, so it's the better chip to use where those features are valued and desired. K20x uses 14 of 16 SMXes, same as the Titan. This is the first K110 announced with this many cores enabled. I'm not sure what configuration the supercomputer customers got, but I'd guess it was probably the 14 SMX version.

Yeah, thats what I meant, first it was the Tesla K20, and then the K20x, (who would possibly need that much GPGPU computing? and then nVidia saw AMD was winning in the GPU wars 6xx vs 7xxx) gutted it and optimized it for gaming, and then the GTX Titan gaming version of it came into being, then the dumbed down slightly yet a much better value GTX 780 which also uses a GK110, and now the "business" version Quadro. Seems like a logical progression, although I haven't seen one benchmark where Quadros beat GeForce's, I've actually seen GeForce's recommended in business situations over Quadros due to their extreme price (on AVADirect a lot, actually...) and here as well.

Why does this Quadro have so much VRAM? What programs use it more than games? AutoCAD? Would Tesla's/Quadros beat GeForce GTX's/Radeons/FirePros in folding? (Since AMD doesn't have a 'Tesla' product division...)

 

[EarthDog]

LOL, games use little vram compared to large AutoCAD renderings... I mean that is a COMPLETE 3D environment compared to, well, not, for gaming.

Quadros are not intended to beat Geforce in gaming. You pay the premium for the quadro for different drivers and technical support among other things. Geforce cards, I thought, were slows in the CAD stuff due to drivers. Perhaps I am mistaking.

 

[JeremyCT]

Two words: certified drivers.

That allows the Quadros to do a lot of accelerations that the GeForce cards don't do. A lot of work goes into validating those drivers for specific tasks and with specific software packages, because a lot sometimes rides on those engineering models that those pros are working on.

Quadro cards get fun stuff like ECC memory, unlocked DP precision performance (if available), and more vRAM to play with. As ED points out, that extra RAM isn't for gaming at all. These are professional cards for professional applications.

Not for gaming, unfortunately. The huge price tag gets you those certified drivers and guaranteed accelerations in certain applications. Whether it's "worth it" is up to the buyer. It depends on what you're doing exactly.

 

[petteyg359]

Would Tesla's/Quadros beat GeForce GTX's/Radeons/FirePros in folding? (Since AMD doesn't have a 'Tesla' product division...)

Just because AMD doesn't have a "Tesla" line doesn't mean their top hardware isn't similar. See the FireGL S10000. Only 6 GB, but 15% more flops

 

[Hippogriff]

AFAIK, Quadro/FirePro cards are designed for rendering/3d modelling/etc.., while Tesla cards are specifically designed for CUDA GPGPU. I doubt a Tesla card performs really well if used for something different.

ATI doesn't have a similar range of cards because CUDA is proprietary and is, more or less, the standard in scientific computing. Matlab, for instance, doesn't even support natively OpenCL.

 

[Starrbuck]

Maybe the Titan II is coming!