NVIDIA Maxwell Speculation Thread

[DSC]

http://www.geeks3d.com/20140208/gpu...and-opencl-info/comment-page-1/#comment-37277

It’s now clearly exposed in NVCUDA.DLL v334.67

push offset D__cuda_arch_8 ; “-D__CUDA_ARCH__=500″
push offset Maxwell ; “Maxwell”
push offset Compute_50 ; “compute_50″

The lab rats at NVIDIA are already playing with a bunch of Maxwells
GM206 GM204 GM200 GM108 GM107

Maxwell confirmed for Compute Capability 5.0.

Also GM20x must be 20nm Maxwells.

 

[Picao84]

http://www.geeks3d.com/20140208/gpu...and-opencl-info/comment-page-1/#comment-37277



Maxwell confirmed for Compute Capability 5.0.

Interesting. I wonder if the GM2xx parts are 20nm,or is a new codename scheme for what should have been GM11x parts. After all, they had GK208 as some sort of refresh of GK107.

 

[Ailuros]

Interesting. I wonder if the GM2xx parts are 20nm,or is a new codename scheme for what should have been GM11x parts. After all, they had GK208 as some sort of refresh of GK107.

You can call it refresh, but it's rather a cut down or lower end verstion of GK107.

 

[pharma]

http://www.geeks3d.com/20140208/gpu...and-opencl-info/comment-page-1/#comment-37277



Maxwell confirmed for Compute Capability 5.0.

Also GM20x must be 20nm Maxwells.

Will be interesting to read more GM20x as it becomes available ... 20nm sounds feasible.

 

[Ailuros]

Considering the bigger Maxwell chip to follow GM107 (in mid summer?) sounds like a GK106 replacement (called GM206) it remains to be seen if it'll be alone or will come along with the performance GM204 SKU.

If correct hurrayyyyyy the top dog will be called in all likeliness "GM200" meaning that we won't have any "it's called GM210 because GM200 was cancelled..." nonsense

 

[Picao84]

You can call it refresh, but it's rather a cut down or lower end verstion of GK107.

Nevertheless, it was used on same SKUs as GK107 AFAIK, such as GT630/GT640, with same or slighter higher performance in most situations where a low end GPU is used for.

http://hexus.net/tech/reviews/graphics/59081-nvidia-gainward-geforce-gt-640-rev-2-gk208/?page=7

Nvidia has rearchitected GT 640 with, now, higher frequencies allied to a leaner architecture. Continuing with codenames, the GT 640 GK208 is a better bet than the GK107 in almost every scenario. Gaming performance is on a par with the recently released and same-priced Radeon HD 7730, though, while substantially improved, GPGPU throughput, via OpenCL, still lags behind comparable AMD cards'.

 

[Ailuros]

Nevertheless, it was used on same SKUs as GK107 AFAIK, such as GT630/GT640, with same or slighter higher performance in most situations where a low end GPU is used for.

Wherever texel fillrate f.e. was less relevant, yes performance was higher.

 

[Picao84]

Wherever texel fillrate f.e. was less relevant, yes performance was higher.

Still, it was used on same SKUs, so they used it as a refresh, and I am therefore correct

 

[RecessionCone]

If correct hurrayyyyyy the top dog will be called in all likeliness "GM200" meaning that we won't have any "it's called GM210 because GM200 was cancelled..." nonsense

Don't worry, we'll still have the "it's called GM200 because GM100 was cancelled..." nonsense.

 

[Ailuros]

Still, it was used on same SKUs, so they used it as a refresh, and I am therefore correct

Oh that's what it was about it; you're always correct and let's move on

Don't worry, we'll still have the "it's called GM200 because GM100 was cancelled..." nonsense.

*kicks sand and frowns*

 

[iMacmatician]

If GM206 is on 20 nm (we can't confidently say one way or the other since we've had process changes with a +10 on the numbering and no process changes with a +100 on the numbering) then I would like to see a die size comparison between GM107 and GM206. Depending on the specs of GM206 they might end up somewhat close in size to each other, although I would think GM206 might be more expensive due to 20 nm.

 

[LordEC911]

If GM206 is on 20 nm (we can't confidently say one way or the other since we've had process changes with a +10 on the numbering and no process changes with a +100 on the numbering) then I would like to see a die size comparison between GM107 and GM206. Depending on the specs of GM206 they might end up somewhat close in size to each other, although I would think GM206 might be more expensive due to 20 nm.

I would hope they would go a bit larger, closer to GK106 size(~220mm2), so they could at least do another 192bit bus.

 

[Picao84]

The real thing, photoshoped or just a non-reference card?



Taken from KiSUAN at http://www.xtremesystems.org/forums...GTX-Titan-Black-Edition-listing-spotted/page3

He also says GM107 has 640 CUDA cores and not 960...

 

[Kaotik]

Non-reference apparently, and not the first time it pops up (though last time it was only in AIB-slide as a mention, no pictures)

 

[Picao84]

nevermind

 

[Blazkowicz]

I would hope they would go a bit larger, closer to GK106 size(~220mm2), so they could at least do another 192bit bus.

Yep that's maybe nice. Nvidia has been doing this since the GF100 generation. Thus they sell a 64bit GPU, a 128bit one, a 192bit, a 256bit and a 384bit. (except there was no GF109)


If they make a 64bit Maxwell (doubtful?) it will be a 20nm one, which comes late as the process is cheaper, mature and the warts have been sorted out.
/edit : duh, there's that geeks3d thing I failed to read, with mention of a GM108 in it. So the above is false. Cheap, low powered GPU out of the door early on 28nm?
This makes GK208 more redundant by the way.

 

[ams]

He also says GM107 has 640 CUDA cores and not 960...

I'm sorry, but that makes no sense at all. For one, this CUDA core amount is not a multiple of 192 and hence would be a significant rearchitect compared to any purported Kepler.M derivative. Second, GTX 650 Ti already has 768 CUDA cores, so it would hardly make sense for GTX 750 Ti to have less.

 

[Blazkowicz]

Nevertheless, it was used on same SKUs as GK107 AFAIK, such as GT630/GT640, with same or slighter higher performance in most situations where a low end GPU is used for.

http://hexus.net/tech/reviews/graphics/59081-nvidia-gainward-geforce-gt-640-rev-2-gk208/?page=7

But it's too expensive, if you get a GTX 650 (twice the bandwith and fillrate basically) or a R7 240 gddr5 you have much better gaming performance for the same price (or about) and the cards have much better availability too.
I'd love to get that GK208, it's nice and tidy, has single slot cards and has all the latest linux drivers/OpenGL features etc. But at that price point I could get a 2TB hard drive, a Firefox smartphone or whatever.

To me it looks like a good workstation card for CUDA devs, CAD and visualisation stuff.

 

[AnarchX]

I'm sorry, but that makes no sense at all. For one, this CUDA core amount is not a multiple of 192 and hence would be a significant rearchitect compared to any purported Kepler.M derivative. Second, GTX 650 Ti already has 768 CUDA cores, so it would hardly make sense for GTX 750 Ti to have less.

Maybe the trashed the super-scalar approach, which was introduced with GF104.


Wrong CUDA core numbers reported by the driver is a known trick, which used NV in the past to hide specifications of new GPUs.

 

[iMacmatician]

If KiSUAN is correct and the ratio of CCs between the 750 and 750 Ti remains the same as what was rumored (768/960 = 0.8), then the 750 would have 512 CCs which would imply a CC/SMX count of some divisor of 128. The "bad" news is that, given no additional power connectors and a ≥ 1 GHz core clock on the 750 Ti (the same assumptions as before), we have a lower bound of "only" 17.1 GFLOPS/W instead of 25.6 GFLOPS/W. Sure, it's not bad compared to GK107 and GK106, but that number isn't higher than those of some existing 28 nm parts.

I'm sorry, but that makes no sense at all. For one, this CUDA core amount is not a multiple of 192 and hence would be a significant rearchitect compared to any purported Kepler.M derivative. Second, GTX 650 Ti already has 768 CUDA cores, so it would hardly make sense for GTX 750 Ti to have less.

Since GM107 is Maxwell, would it be surprising that it has a different CC/SMX count than Kepler?