GM107 Features 128 CUDA Cores Per Streaming Multiprocessor

[btarunr]

NVIDIA's upcoming GM107 GPU, the first to be based on its next-generation "Maxwell" GPU architecture, reportedly features a different arrangement of CUDA cores and streaming multiprocessors to those typically associated with "Kepler," although the component hierarchy is similar. The chip reportedly features five streaming multiprocessors, highly integrated computation subunits of the GPU. NVIDIA is referring to these parts as "streaming multiprocessor (Maxwell)," or SMMs.

Further, each streaming multiprocessor features 128 CUDA cores, and not the 192 CUDA cores found in SMX units of "Kepler" GPUs. If true, GM107 features 640 CUDA cores, all of which will be enabled on the GeForce GTX 750 Ti. If NVIDIA is carving out the GTX 750 by disabling one of those streaming multiprocessors, its CUDA core count works out to be 512. NVIDIA will apparently build two GPUs on the existing 28 nm process, the GM107, and the smaller GM108; and three higher performing chips on the next-generation 20 nm process, the GM206, the GM204, and the GM200. The three, as you might have figured out, succeed the GK106, GK104, and GK110, respectively.



View at TechPowerUp Main Site

 

[Arjai]

So, basically the only thing going for these is DX11.2. Is that enough?

 

[HalfAHertz]

Well...they are less but judging from the preliminary results, more powerful than the old SMs, so in the long run it should equal out? I guess that's how NV got those better power numbers - increased complexity but reduced the overall number of SMs.

Oh and GPUz was wrong

 

[Patriot]

Well...they are less but judging from the preliminary results, more powerful than the old SMs, so in the long run it should equal out? I guess that's how NV got those better power numbers - increased complexity but reduced the overall number of SMs.

Probably more high level tuning than retooling the individual SP.
Balancing the components for efficiency gains, eliminating bottlenecks.

 

[Nordic]

So stream processors?

 

[Patriot]

So stream processors?

SP = cuda cores...
They are saying there are less SP/cudacores in each CU compute unit or SMM

 

[renz496]

So, basically the only thing going for these is DX11.2. Is that enough?

isn't that current kepler is DX11.2 capable as well?

 

[Patriot]

So, basically the only thing going for these is DX11.2. Is that enough?

isn't that current kepler is DX11.2 capable as well?

no... not even 11.1...
Software yes... but not hardware.

They support some of the features, but they are not DX 11.1 or 11.2 compliant.

Efficiency gains is where maxwell is supposed to be dominating on.

 

[Xzibit]

isn't that current kepler is DX11.2 capable as well?

Nvidia site has all the 600 & 700 series listed as only 11

Even the latest GTX 780 Ti is only 11 compliant

Microsoft® DirectX® 11.2 (feature 11_0)

DirectX 11.2 API (feature level 11_0)

 

[esrever]

With this configuration, they can do 1/2 rate DP again which is how they are going to increase performance per watt they were promising.

 

[Arjai]

So, somewhere, I read an article, linked in another 750 post, that said these 'new' cards were 11.2...basically, everything about these things is conjecture.

Perhaps, on the 18th, the conjecture will end? 6 more days. Have fun! I'm gonna wait and hopefully see some real tests, with real cards.

 

[haswrong]

but but but, will it support mantle?

 

[Fluffmeister]

but but but, will it support mantle?

Highly unlikey, nVidia users like to use decent CPU's so no need.

 

[Ghost]

but but but, will it support mantle?

Yes, it will. But only when Radeons start supporting CUDA along with PhysX.

 

[SIGSEGV]

Highly unlikey, nVidia users like to use decent CPU's so no need.

lmfao..
dude.. i must admit it.. that you're very funny..

 

[semantics]

Nvidia site has all the 600 & 700 series listed as only 11

Even the latest GTX 780 Ti is only 11 compliant

Somewhat sure they support DX11.1 features but only the ones that nvidia felt were perternant to games so they didn't get official compliance.

 

[Fluffmeister]

Somewhat sure they support DX11.1 features but only the ones that nvidia felt were perternant to games so they didn't get official compliance.

Yeah It's basically a non issue, but peeps like to argue any advantage that they can.

We have been inundated with titles using DX11.2 after all.

 

[badtaylorx]

i hate to say this,

but maxwell is looking pretty week....

 

[lemonadesoda]

I guess with experience, they can now optimise the processor for the typical GPU and CUDA workload. If the device is cheaper, more efficient, cooler, I'm ok with it

 

[WhoDecidedThat]

Thank God. They are moving away from superscalar processing.

Supercalar Processing: http://www.anandtech.com/show/3809/nvidias-geforce-gtx-460-the-200-king/2

 

[Fluffmeister]

i hate to say this,

but maxwell is looking pretty week....

I hate to say this, but it will be looking pretty next week.

But seriously, it's a GX107 part, of course it's not going to set the world alight on pure performance alone, but it's how it reaches that performance that will be interesting, especially how it compares to Kepler on the same process will be rather telling.

 

[xorbe]

maxwell is looking pretty week....

It's hanging in there with 640 cores / 128 bit bus, and it looks weak? It's a good sign for the beefier versions, I figured.

 

[BiggieShady]

i hate to say this,

but maxwell is looking pretty week....

I wouldn't judge their future big chip (GM200) based on small one (GM107).
Couple of things are interesting compared to Kepler : increased number of instructions per cycle, bigger cache and 15% larger transistor density on the same 28nm node.

 

[TheinsanegamerN]

I cant wait to see what maxwell will be able to accomplish on a 20nm node with, say, the 180 watt TDP of a geforce 770. geforce 870 at 180 watt with performance of oced 780ti?

 

[Xzibit]

I wouldn't judge their future big chip (GM200) based on small one (GM107).
Couple of things are interesting compared to Kepler : increased number of instructions per cycle, bigger cache and 15% larger transistor density on the same 28nm node.

No kidding it has more L2 cache than the TITAN/780 Ti

GM107 = 2mb
GK107 = 256kb
GK110 = 1.5mb
Hawaii = 1mb

12mb of L2 cache on die soon...