Nvidia BigK GK110 Kepler Speculation Thread

[lanek]

A claimed GPU-Z of the GTX 780 Ti from Chiphell:

The GFLOPS (4503) are very close to Titan's (4500). Also, MyDrivers.com has a rumor of a "new GTX 780" that is not the 780 Ti.

Why not, but i dont see the point of release another 780 " non TI " .. something between 770 -780 maybe...

 

[trinibwoy]

Pretty lame if true.

 

[AnarchX]

The fill-rates in this GPU-Z shot are wrong. But of course GPU-Z has sometimes problems with calculation.

Also, MyDrivers.com has a rumor of a "new GTX 780" that is not the 780 Ti.

Maybe this:

NVIDIA GeForce GTX 780 Rev. 2
[...]
Bios Remade from the ground up

http://www.techpowerup.com/gpudb/2513/geforce-gtx-780-rev-2.html

 

[lanek]

For what i read on other forums, the GPU-z screenshot was a fake, someone who play with infos on GPU-z was just do an example of what it can be. ( i dont know enough about the story for be sure ).

 

[iMacmatician]

According to this document from HP, the Quadro K6000 uses Boost for a base clock of 797 MHz and a boost clock of the 902 MHz that we are familiar with.

I should have known a 902 MHz base clock in 225 W is too good to be true. Still, that's a nontrivial bump over the K20X.

 

[mczak]

According to this document from HP, the Quadro K6000 uses Boost for a base clock of 797 MHz and a boost clock of the 902 MHz that we are familiar with.

I should have known a 902 MHz base clock in 225 W is too good to be true. Still, that's a nontrivial bump over the K20X.

9% higher core (base) clock, 15% higher memory clock and lower TDP (225W vs. 235W) compared to K20X. That's quite ok for something which is supposedly more or less the same chip.

 

[boxleitnerb]

And 1 SMX more. But somehow I doubt the 780 Ti will have all 2880 SPs enabled...

 

[rapso]

the thing pointed out at the NV conference was "more power efficient". I did not hear "faster" or something else regarding the TI. might be the same configuration but with less power drawing. boost might last longer in benchmarks, but effectively same speed for games.

 

[mczak]

And 1 SMX more.

Ah yes forgot about that. Sounds really good then (17% more compute power, 15% more memory bandwidth and all with a 5% reduction in TDP so about a 20% power efficiency increase) though I'd want to see real-world power draw measurements. It also has twice the memory but since it's just 4gbit instead of 2gbit chips that shouldn't really make a difference in power draw.

 

[Psycho]

Considering how long it took from the first K20 shipments and until the chip was mature enough for the consumer market those early specs have probably been pretty conservative.
And I'm not sure we can really compare server and desktop tdp values either - the former is much more likely to be taken at face value.. (even a "power virus" should neither throttle or exceed the tdp)

 

[AnarchX]

GK180 = GK110-B1?

The later one is already shipped on OC GTX 780:

http://www.xtremesystems.org/forums...Y-GTX-780-Hall-of-Fame-(HOF)-Performance-Test

TPU has the B1 added in its database, too:
http://www.techpowerup.com/gpudb/2513/geforce-gtx-780-rev-2.html

 

[lanek]

We are in the black zone here AnarchX, i dont know. TDP wise it is really hard to compare a Geforce and Quadro-Tesla, the qualification are absolutely not do the same way.

 

[iMacmatician]

From Fudzilla: "Geforce GTX 780 Ti to end up faster than Titan."

We are not sure if GTX 780 Ti beats the Titan in all benchmarks, but it will definitely be faster in most of them.

If the 780 Ti has equal or better specs than the Titan in every area (including the VRAM size and boost headroom) then I don't see why it would ever perform worse than the Titan, so if there is some benchmark where the 780 Ti loses, then that suggests some tradeoff in specs from the Titan to the 780 Ti (even if it's an overall upgrade). They don't say what benchmarks that they are not sure if the 780 Ti will win against the Titan in though (DP? Ones where if the 780 Ti were 3 GB it would be VRAM-limited?).

 

[xDxD]

[XC] Oj101 on XtremeSystem:

"Stock clock speed on the 780 Ti puts it between 3 and 5 % below a 290X running at around 1150 MHz on the Performance Preset.

By the way, I'm pretty sure the 780 Ti will use a new revision of GK110, not GK180. I don't see any need for validating ECC, etc on a consumer desktop card. GK110 can indeed have more CUDA cores."

 

[A1xLLcqAgt0qc2RyMz0y]

[XC] Oj101 on XtremeSystem:By the way, I'm pretty sure the 780 Ti will use a new revision of GK110, not GK180.

If the GK180 is a respin with better yields than the GK110 at higher clocks and fully enabled then it more likely will be used for the 780 Ti.

I don't see any need for validating ECC, etc on a consumer desktop card. GK110 can indeed have more CUDA cores."

Why would any validation of ECC need to be done at all on a consumer GPU.

 

[iMacmatician]

The 2013 IEEE conference on High Performance Computing (December 18-21) has a session titled "NVIDIA Tesla K40 and CUDA 6.0 Launch in India."

This NVIDIA session is scheduled for the 19th (according to the Advance Program page).

 

[trinibwoy]

Does anyone have hard info on Kepler's register file bandwidth? The maximum throughput I'm seeing on a 680 is 128 instr/clk per SM using gpubench. Same goes for these guys - http://hal.inria.fr/docs/00/78/99/58/PDF/112_Lai.pdf.

Sorry for the self-quote but this is still a mystery to me. It's pretty clear at this point that a Kepler SMX can only issue 128 ALU instr/clk peak, regardless of what NVidia claims.

Does anyone have any idea why they would have 6 SIMDs per SMX when only 4 warps can be issued to them? Do they have some sort of CPU like setup where instruction issue ports are shared by the execution units and having more SIMDs helps to resolve port conflicts?

I'm sure they serve some purpose - just wish I knew what it was!

ftp://ftp.lal.in2p3.fr/pub/PetaQCD/talks-ANR-Final/Review_Junjie_LAI.pdf

 

[fellix]

Is this valid for GK110, too?

 

[Jawed]

OK, my mind is blown. This would appear to imply that Titan is at about the same SGEMM performance as HD5870 (nearly 2 TFLOPS).

DGEMM should be fine though (should be 90%+ of theoretical).

 

[RecessionCone]

Is this valid for GK110, too?

Yes.