NVIDIA Maxwell Speculation Thread

"Gigabyte has produced a Mini-ITX GeForce GTX 970 Graphics Card, yeah it's cute alright. You soon will be able to find it under SKU code GV-N970IXOC-4GD. The card is running at 1076 MHz Base with a Boost clock of 1216 MHz, which is actually a notch faster then the reference cards. The card has a single 8-pin PCI-Express power connector."

Ready for your Steam Machine XD
 
Looks neat :)

index.php
 
Now I want one for my media center :oops:

And here I thought I'd never upgrade the media center again because of Steam Home Streaming.. but that thing seems to have pretty much stalled for the past 6 months..
 
Kepler and Maxwell have allowed NV to make some seriously powerful little cards. Even the normal GTX 970 is pretty short and makes its dual PCIe power sockets seem out of place. My 560 Ti and 6950 are considerably longer cards than the 970 that I just got.
 
Kepler and Maxwell have allowed NV to make some seriously powerful little cards. Even the normal GTX 970 is pretty short and makes its dual PCIe power sockets seem out of place. My 560 Ti and 6950 are considerably longer cards than the 970 that I just got.

Yeah I don't have a very large case and the MSI Gaming model fit perfectly. The Gigabyte models are insanely large in comparison, but they have 3 fans of course. 15.6" for the G1 model! :oops:
 
Yeah I don't have a very large case and the MSI Gaming model fit perfectly. The Gigabyte models are insanely large in comparison, but they have 3 fans of course. 15.6" for the G1 model! :oops:

Gigabyte G1s are 12.2 inches long.

Have to try that DSR with my 780TIs.
 

Newegg has taken that page down (your link is fine), possibly they are correcting their mistake.

I believed you didn't come up with those numbers in your head anyway :) I have to say though that especially the height should be a dead giveaway that those specs are incorrect. It's still a big card at 12.2". 15.9" x 9.3" would be 40.3cm L x 23cm H, (Standard DVD box is 19cm high.)

index.php


http://www.guru3d.com/articles_pages/gigabyte_geforce_gtx_970_g1_gaming_review,3.html
 
Huh that's funny. Thanks for the clarification! And yes it's still a big card at a foot in length!
 
GM206 launch?

Any updates on GM206's launch? It was supposed to follow GM204 very closely and so far we've barely seen any leaks.

PS: My own personal speculation is 2xGM107, so a 10 SMM part with a 192 bit MC.
 
GM206 samples were much later at Zauba, early September. The rumored Q1 2015 launch window for GTX 960 would match this.
There are rumors in Asia that GM206 could be only a 8 SMM part (this would be comparable to GF106/116:GF104).
 
GM206 samples were much later at Zauba, early September. The rumored Q1 2015 launch window for GTX 960 would match this.
There are rumors in Asia that GM206 could be only a 8 SMM part (this would be comparable to GF106/116:GF104).

Hmm then I guess there has been some delay..or maybe it needed a respin because it taped out not too long after GM204 did.

Yes but GF106 was also pretty much 2xGF108, which is why I was speculating 2xGM107. 8 SMM's would be too close to GM107 IMHO.

Also..I am curious as to what they would do with the L2. L2 has always been mapped to the MC's. So with a 192 bit MC there would be no easy way to get 2MB L2 like GM107 and GM204 have.
 
How does maxwell clock so high in the first place?

Much of the superiority of gm204 is related to its high clocks even with its increased IPC.

Nvidia increased IPC(this I get from Ryan's explanation at ANAND)and increased its clock by 25%(how did they manage this?) and somehow, it didn't increase power consumption.

The increased IPC usually increases power consumption like it did with Hawaii. It also lowers clocks(again like hawaii or Conroe). How did Nvidia get the best of both worlds without a die shrink?
 
Also..I am curious as to what they would do with the L2. L2 has always been mapped to the MC's. So with a 192 bit MC there would be no easy way to get 2MB L2 like GM107 and GM204 have.
Why must be there 2MiB? GM107 big L2 is probably based on the DDR3 SKUs. So GM206 192-Bit should be 1.5MiB.

GM107 could be a dead end, since it is Maxwell 1.0 and probably no DX12. A close to 200mm² DX12 part could be a great seller.
 
A little bit more on VXGI (voxel processing):

Jon Peddie ...

Nvidia thinks gamers currently using 780-class graphics boards are longing for an upgrade. But the big deal will be voxel processing. The Maxwell GPU will implement Nvidia’s VXGI real-time global illumination capability. It’s capable of casting a configurable number of cones (probably 5 to 15 per pixel, depending on the performance of the GPU) to trace through the voxel grid and collect indirect light to apply per pixel.

VXGI can eliminate the need for pre-illumination lighting. With voxel illumination, it is possible to get specular and reflective light. Nvidia expects game engines to incorporate its voxel indirect illumination capability by the end of the year.

Nvidia is first to use a GPU to do voxel global illumination, but not the first to use voxels. Certainly they won’t be the last. AMD will likely offer a similar capability. Nvidia spoke about voxel cone processing at Siggraph 2011 and, earlier this year, at GTC, and surely AMD was listening.

Nonetheless, Nvidia has taken the lead, injecting voxels back into our vocabulary, and it has gotten the cooperation of the major game engine developers to incorporate such capabilities. The game developers will use OpenCL rather than Nvidia’s Cuda so they can run on other processors, too.

Over time other processor suppliers will vaunt their voxel-ness. But just as it has done with GPU compute and workstation graphics, Nvidia has taken and probably will hold the lead dog position in voxel land, too.
http://www.eetimes.com/author.asp?section_id=36&doc_id=1323973
 
How did Nvidia get the best of both worlds without a die shrink?

Well as per Anandtech, due to the maturity of the 28nm process, Nvidia also implemented some low level transistor level optimizations for increased energy efficiency. Nvidia obviously wasn't very open about this and so it is difficult to quantify the effect of this.

And of course the biggest effect was due to the architectural changes. I am cross posting silent guy's excellent post on this from the Tegra thread - http://forum.beyond3d.com/showpost.php?p=1883353&postcount=3100
- simplification of operand fetching from the register bank: no more power hungry crossbar to swizzle registers around. (See SGEMM article of Maxas.)
- A register cache to reduce the number of register bank accesses and reduce the distance the operands need to travel. (Same SGEMM article.) This is because, for layout reasons, large RAMs always need to be placed on the border of a block.
- Reduced number of ALUs where operations can be scheduled which reduces the complexity of the routing logic.
But there are more:
- operation scheduling removed from HW and moved to the compiler. (Slides from Nvidia/Confirmed by Maxas.)
- Larger register files which increases the number of registers that can be used in a shader. This improves occupancy and reduces the number of power costly round trips to the L2 cache.
- Increased utilization of the available ALUs: SMX of Kepler has 192 ALUs but even Nvidia's SGEMM library never manages to exceed 75(?)% utilization. SMM has 128 ALUs and the Maxas guy gets to something like 96%.
Reduced architectural inefficiency, one might be tempted to speculate.

Interesting phrasing. Wouldn't it have been simpler for one to say "Increased architectural efficiency" instead? :D
Why must be there 2MiB? GM107 big L2 is probably based on the DDR3 SKUs. So GM206 192-Bit should be 1.5MiB.

GM107 could be a dead end, since it is Maxwell 1.0 and probably no DX12. A close to 200mm² DX12 part could be a great seller.
I'm not saying there must be..was just speculating that it could be, based on the 2MB present in GM107, and then subsequently on GM204. Possibly it may be a minimum design criteria going forward. Going to 1.5 MB in GM206 would be a bit of a regression if this was the case.

GM107 is still quite a good chip, I dont see it going anywhere, anytime soon, especially for laptops.
 
Back
Top