You cannot share L1 without sharing SMEM (or local memory in your tongue) aswell. They are the same reconfigurable cache attached to the same set of load store units on the chip.
It is true however that each sub-unit would need their own set of LDS.
There is a lot more in Kepler's SMX tied in a fixed manner than one could think from glacing over the usual block diagrams.
WRT to local memory/shared memory an L1 think about this: Schedulers own parts of the register files in Kepler as well as specific functional units. Now there's 64 kiB RAM, configurable. What's the difference between L1 and shared mem globally speaking? Only for Shared Memory do you allow other warps/schedulers/sets of functional units to access this part of the memory. Since this needs to be set beforehand, you can software-configure it. I see absolutey no problem in what Novum proposes - in fact I think that this is the case with Kepler as well already.
WRT to local memory/shared memory an L1 think about this: Schedulers own parts of the register files in Kepler as well as specific functional units. Now there's 64 kiB RAM, configurable. What's the difference between L1 and shared mem globally speaking? Only for Shared Memory do you allow other warps/schedulers/sets of functional units to access this part of the memory. Since this needs to be set beforehand, you can software-configure it. I see absolutey no problem in what Novum proposes - in fact I think that this is the case with Kepler as well already.
I heard that the four subdivision will share one 64KB LDS, and one L1Data/texture cache plus the texturing block. The LDS bandwith is 64 byte/cycle. Each subdivision has a 64KB register file and 8 LD/ST units.
But I don't have the cards, I just speak with my Chinese friend.
It won't be a compute monster if true. Maybe even worse than Kepler.
It's a disappointment if it doesn't come with more SM/ALU.
That's an implementation detail which can change.
iMacmatician
Regular
3 SMM and 64-Bit DDR3 (maybe also GDDR5), according to some Asian leaks.
To sad, there is probably no GM106/105 (1280CC, 256-Bit), which could be a nice ~250mm²/$250 solution.
When should the reviews be out?
I'm really curious on the impact that big cache could have on performances
Supposedly next Tuesday, 18th February.
Nvidia already has GM20x in their drivers, they're just waiting for TSMC to start production. Since all Nvidia architectures are now designed for mobile first & power efficiency and going bottom to top with GM107 first release, GM206 should be the next release, followed by GM204 and finally GM200. There's no point in making 28nm GM106 etc.
iMacmatician
Regular
I must have been unclear, I was wondering if the GM108 shader count was also fixed.
Nvidia brought GK107 to market (in notebooks) before GK104. GK104 came next, followed by GK106. So it's entirely plausible GM104/204 will come next.
If 20nm is still a ways off, or if Nvidia is planning to just skip 20nm and go straight to finfets, then we will surely get GM104/204 on 28nm. Considering the perf/watt and perf/mm^2 improvements GM107 has over GK107, the successor to GK104 could happily exist on 28nm if production costs make sense.
20SoC should be feasable starting from H2 this year, however I've no clue how costs would look like there. TSMC started mass production for 20SoC recently didn't they? They plan to start with 16 FF mass production in early 2015; what makes anyone think that 16 FF will be at the same time affordable in early 2015?
Yes GK107 was in OEMs hands in December 2012 and GK104 would had appeared earlier if they wouldn't had decided to give it a minor boost to go head to head with Tahiti.
Now the Maxwell family is a couple of months later and while GK110 had its final tape out if memory serves well in early March unless I've missed something the newest succeeding egg hasn't hatched yet.
Yes GK107 was in OEMs hands in December 2012 and GK104 would had appeared earlier if they wouldn't had decided to give it a minor boost to go head to head with Tahiti.
Now the Maxwell family is a couple of months later and while GK110 had its final tape out if memory serves well in early March unless I've missed something the newest succeeding egg hasn't hatched yet.
Blazkowicz
Legend
I agree that GM204 comes next, this way nvidia may get the "gaming flagship" out and still sell the GK104 desktop cards below it.
Laptops can use the GM204, low-clocked and cut down if need be, esp. in "gaming laptops". GM206 laptops will come later anyway.
Laptops can use the GM204, low-clocked and cut down if need be, esp. in "gaming laptops". GM206 laptops will come later anyway.
http://www.3dcenter.org/news/chipna...liche-spezifikationen-der-maxwell-grafikchips
guys what do you think of these speculations? would be realistic a GM200 and GM204 like those?
When you're using such a big range it's hard to not hit at least a couple of them
Not really
Not really
