Nvidia BigK GK110 Kepler Speculation Thread

[xDxD]

The upcoming Tesla K20 card will be equipped with a whopping 2880 Shader / ALU processor encounting monster.
Update: new information shows that there actually might be 16 SMX cores and thus 3072 Shader processors on GK110.


http://www.guru3d.com/news/gk110based-surfaces--has-2880-shader-processors/

 

[3dilettante]

We do have a picture of the chip for the sake of an SMX count.
Not sure where to find the 16th.

 

[Lightman]

We do have a picture of the chip for the sake of an SMX count.
Not sure where to find the 16th.

Maybe nVidia took a page of AMD and they showed only artists impression of GK110 die?

 

[CarstenS]

New stuff from Inside Kepler Session:
http://www.pcgameshardware.de/aid,8...-GTC-2012-vorgestellt-Kopie/Grafikkarte/News/

7 SMX, 1.5 MiB L2, 255 registers per thread adressable, 4x 16 FP64-Units per SMX according to block diagram and a separate path from texture cache to SMX, 66% lower ECC overhead than on Fermi

 

[RecessionCone]

Has a white paper been released for the GK110?

http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

 

[Albuquerque]

Our rendering servers delivers the highest throughput with 2 instances (of our application, ie 1 d3d9 context pr instance) per graphics card, and adding another card with it's own load can barely affects performance (ie it scales nicely).
This is Windows7 (for the particular machine with 2 cards, but 2008R2 works the same) and radeons.

I know I'm late to the party, and this was posted a while back. Can you share any more info on your app? I've been doing some tinkering with Microsoft Hyper-V and RemoteFX platforms, specifically around the "shared" 3D acceleration that RemoteFX can deliver across multiple simultaneous streams via any 3D accelerator card that has a WDDM driver available. Even an ATI 4850 provides an acceptable amount of performance for 'easy' games across three separate Win7 VM remote sessions.

I'm not entirely sure of how they do it, either.

 

[trinibwoy]

http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

Thanks!

 

[Tim]

http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

Interesting:

"SMX: 192 single‐precision CUDA cores, 64 double‐precision units, 32 special function units (SFU), and 32 load/store units (LD/ST)."

 

[fellix]

So... the texture cache in Kepler now has direct read-path to the SIMD lanes? Quite useful.

 

[ninelven]

So...
1/3rd rate DP
Access to texture cache
2x L2
4x Registers
2880 CCs
240 TUs

I wonder if it will be substantially faster than GK104 in games?

 

[fellix]

Another interesting bit:

Unlike Fermi, which did not permit double precision instructions to be paired with other instructions, Kepler GK110 allows double precision instructions to be paired with certain other instructions that have no register file reads, including load/store, texture, and some integer instructions.

 

[iMacmatician]

So...
1/3rd rate DP

>1 TFLOPS of 1/3 DP at ~85% efficiency gives >3.53 TFLOPS SP, which indicates >613 MHz with 15 SMs, >657 MHz with 14 SMs, or >707 MHz with 13 SMs.

Interesting those hypothetical clocks are so low….

 

[trinibwoy]

I wonder if it will be substantially faster than GK104 in games?

Depends on clocks but it's almost guaranteed to be worse perf/w unless nVidia found additional architecture tweaks. Too bad they didn't announce a release date for consumer versions.

 

[mczak]

Depends on clocks but it's almost guaranteed to be worse perf/w unless nVidia found additional architecture tweaks. Too bad they didn't announce a release date for consumer versions.

I'm not sure it will be worse perf/w. Sure at the same clocks it would be. But chances are clocks (and hence more important voltage) will be lower which can improve things a lot. So might be something like 50% faster with 50% higher power draw in the end.

 

[Man from Atlantis]

New stuff from Inside Kepler Session:
http://www.pcgameshardware.de/aid,8...-GTC-2012-vorgestellt-Kopie/Grafikkarte/News/
Shuffle-HW-Instruktion - Atom-HW-Instruktion - Bit-Shift, Bit Rotate, FP32 division und read-only cache über den Compiler

This thread was lacking bitcoin talk but apparently they should be happier now

 

[Man from Atlantis]

Another interesting bit:

Does that mean DP units can be used as additional SP units? 2880 + 960(or even x2)

 

[Gipsel]

Does that mean DP units can be used as additional SP units? 2880 + 960(or even x2)

No. First, nV would have told us so and would advertize the higher speed. They explicitly say, that the co-issued instructions (to a DP instruction) cannot read from the register file. I imagine there can't be too many instructions of this type.
And as a second point: the actual implementation will probably look differently than what nvidia sketched.

 

[denev2004]

I still can't understand the benefits of designing SP/DP unit seperately..

 

[ninelven]

I doubt it is really separate...

 

[UniversalTruth]

I wonder if it will be substantially faster than GK104 in games?

Well, unfortunately they will release GeForces with GK110 only in 2013.

GK110 will not appear in Quadro or GeForce products this year. It is expected that NVIDIA will fight AMD’s new Graphics Core Next based FirePro at SIGGRAPH 2012 (early August 2012), with GK107 and GK104 based parts, while professionals will probably have to wait for GK110-based Quadro until April 2013 (NAB 2013). GeForce gamers should get their hands on GeForce GTX 700 Series sooner, though. Did we just mention GTX 780? Dang...

Read more: http://vr-zone.com/articles/nvidia-...rce-and-quadro-cards/15884.html#ixzz1v6ePIBpA

But, does it mean we will see GK114 as GTX 760 while a little bit later GK110 as GTX 780?

2880 CUDA Cores
15 SMX Clusters
384-bit Memory Controller
Up to 24GB of GDDR5 memory
2nd Gen ECC
Hardware GPU Silicon Virtualization
Hyper-Q (Slashes CPU idle time by allowing multiple CPU cores to simultaneously utilize a single Kepler GPU, dramatically advancing programmability and efficiency)
Dynamic Parallelism (Simplifies GPU programming by allowing programmers to easily accelerate all parallel nested loops – resulting in a GPU dynamically spawning new threads on its own without going back to the CPU)
50-85% Double Precision Rate to Single Precision
At least 1.5 TFLOPS DP FP64
Target: 250 GB/s bandwidth