Nvidia BigK GK110 Kepler Speculation Thread

[UniversalTruth]

"Binned for efficiency" is a relative term. It's still a hugely power-hungry GPU and the available space for air intake doesn't get bigger just because the GPUs are downclocked a bit.

It is a three-slot card by reference design which would supposedly mean it has quite a lot of metal surrounding the chip for cooling too.

 

[AnarchX]

GK210 shipping? There seems to be some compatibility between GK110, GK180 (dev-step to GK210?) and GK210:

https://www.zauba.com/import-gk210-hs-code.html
https://www.zauba.com/import-gk180-hs-code.html

 

[Picao84]

GK210 shipping? There seems to be some compatibility between GK110, GK180 (dev-step to GK210?) and GK210:

https://www.zauba.com/import-gk210-hs-code.html
https://www.zauba.com/import-gk180-hs-code.html

What should a GK210 be exactly

 

[AnarchX]

What should a GK210 be exactly

http://forum.beyond3d.com/showpost.php?p=1827639&postcount=937

Maybe some kind of answer to Hawaii with its 2,6TFLOPs DP @ FirePro. With GK2xx roots we can speculate, that TMU count is halfed like on GK208 and GM1xx. This could saved some transistors to go half-rate DP too.

The "GK210-885-A1" should be the GPU string of a Quadro SKU, which goes on sale. Qualification/engineering GPUs have other strings.

 

[Alexko]

http://forum.beyond3d.com/showpost.php?p=1827639&postcount=937

Maybe some kind of answer to Hawaii with its 2,6TFLOPs DP @ FirePro. With GK2xx roots we can speculate, that TMU count is halfed like on GK208 and GM1xx. This could saved some transistors to go half-rate DP too.

The "GK210-885-A1" should be the GPU string of a Quadro SKU, which goes on sale. Qualification/engineering GPUs have other strings.

That sounds like a lot of effort just to get a bit more DP performance, not to mention the overlap with Maxwell. If there really is a GK210, I would bet on a respin of GK110 (or GK118 as the latest version is apparently called). Plus Hawaii came out five months ago, that's far too little time to design a new chip.

 

[Picao84]

That sounds like a lot of effort just to get a bit more DP performance, not to mention the overlap with Maxwell. If there really is a GK210, I would bet on a respin of GK110 (or GK118 as the latest version is apparently called). Plus Hawaii came out five months ago, that's far too little time to design a new chip.

But wouldn't a change from 1 to 2 on first number imply more than a respin?

 

[AnarchX]

SM_37 and the shared memory increase to 80KiB is a bit to much for just a respin. And NV should had a good picture of Hawaii in end of 2012/early 2013.

 

[Blazkowicz]

Isn't it just Tegra?, does Tegra K1 have feature level 3.5 or 3.7?

 

[Picao84]

Isn't it just Tegra?, does Tegra K1 have feature level 3.5 or 3.7?

Since those links point to a GPU being shipped somewhere and not to a SoC, I figure not.

 

[Alexko]

But wouldn't a change from 1 to 2 on first number imply more than a respin?

Maybe, but really, who knows? There's not much precedent for this. Tesla somehow started at GT2XX, Fermi started and stopped at GF1XX, and so far the only GK2XX chip is GK20A, i.e. Tegra.

 

[Kaotik]

But wouldn't a change from 1 to 2 on first number imply more than a respin?

Metal layer changes would change the last number, the x in -Ax.
Full respin would change the A to B to C etc
210 would indicate completely new chip

 

[iMacmatician]

http://forum.beyond3d.com/showpost.php?p=1827639&postcount=937

Maybe some kind of answer to Hawaii with its 2,6TFLOPs DP @ FirePro. With GK2xx roots we can speculate, that TMU count is halfed like on GK208 and GM1xx. This could saved some transistors to go half-rate DP too.

The "GK210-885-A1" should be the GPU string of a Quadro SKU, which goes on sale. Qualification/engineering GPUs have other strings.

All the Kepler GPUs so far have a 1/(3·2^n) DP rate. Is it possible to go 1/2? I wonder if 2/3 is possible (due to the separate DP units). That would give 3.1 DP TFLOPS for a GK210 with the same number of SP units and the same clocks as the K6000.

Maybe, but really, who knows? There's not much precedent for this. Tesla somehow started at GT2XX, Fermi started and stopped at GF1XX, and so far the only GK2XX chip is GK20A, i.e. Tegra.

There's also GK208.

 

[RecessionCone]

Maybe, but really, who knows? There's not much precedent for this. Tesla somehow started at GT2XX, Fermi started and stopped at GF1XX, and so far the only GK2XX chip is GK20A, i.e. Tegra.

G80 was also known as GT100.

 

[Kaotik]

Maybe, but really, who knows? There's not much precedent for this. Tesla somehow started at GT2XX, Fermi started and stopped at GF1XX, and so far the only GK2XX chip is GK20A, i.e. Tegra.

You forgot GK208

 

[mczak]

All the Kepler GPUs so far have a 1/(3·2^n) DP rate. Is it possible to go 1/2? I wonder if 2/3 is possible (due to the separate DP units). That would give 3.1 DP TFLOPS for a GK210 with the same number of SP units and the same clocks as the K6000.

I don't think 1/2 would work easily with Kepler. Unless your Kepler chip has a Maxwell-style SMM . 2/3 can't neither (you would not have enough register fetch bandwidth).

 

[Ailuros]

I don't think 1/2 would work easily with Kepler. Unless your Kepler chip has a Maxwell-style SMM . 2/3 can't neither (you would not have enough register fetch bandwidth).

Wait a second I thought GK110 has 64 FP64 SPs/SMX; IF they would had gone through lengths to implement 96 FP64 SPs/SMX in a followup Kepler chip they would had also catered for everything else also. I just don't see them developing a separate chip just for the HPC market.

 

[Alexko]

You forgot GK208

Yes, my bad. And GK20A is more or less half a GK208.

 

[Blazkowicz]

You forgot GK208

You forgot Poland!

 

[Picao84]

You forgot Poland!

Sometimes I wonder if you post when you are drunk

 

[mczak]

Wait a second I thought GK110 has 64 FP64 SPs/SMX; IF they would had gone through lengths to implement 96 FP64 SPs/SMX in a followup Kepler chip they would had also catered for everything else also. I just don't see them developing a separate chip just for the HPC market.

You are right 1/2 FP64 rate probably wouldn't really be all that difficult to implement in some tweaked Kepler chip. I was mostly thinking about how the DP units are arranged within a SMX (based on the neat diagrams, http://www.hardware.fr/articles/916-3/maxwell-1st-gen-smm-details.html) but thinking about it nothing would really exclude having DP units in the shared part.