Not much different in what we can see in the GK104 die-shot, with the difference that GK110 now doubles the L2 size per partition, incl. ECC bits and logic and probably some additional HPC-related RAS features.
What do you mean? I thought all of the are surprisingly similar?
Simple logic cells are aligned in the same direction, irrespective of the axis orientation of the whole block. That's why there are visible layout differences. You can see the same thing on any hi-res AMD CPU die-shot, for instance.
And the available space is simply a bit differently shaped. So it is not just a rotated copy as with the (mirrored) SMXs, they did a slightly different layout for the two versions to acommodate for that.
iMacmatician
Regular
I will not say a typo, but can be just the name of the sku used as the project was running. The next after this refresh will be Maxwell so surely something like GM 100
If we expect the logical move to GK11x plateform as Nvidia have do by the past for the next one.. this GK208 could just be a low end card ( maybe even a GK107 card ) declassed for future low end graphic market OEM..
Blazkowicz
Legend
No, those are really chip codenames. A crippled GK107 card would be GK107. (The Quadro 410 is one such card, did it have a very low profile launch? I found it in retailer websites but nvidia didn't boast about it, Quadro 600 and higher Quadros have to be sold)
Kepler has a refresh, whereas Fermi didn't really have one, it was the same chip with reworked transistors. So GF10x -> GF11x wasn't a big step. (though GF104 vs GF100 could be called a slight refresh)
Here GK104 -> GK110 is a bigger step, but it's the same basic thing with features that were omitted on GK10x.
They could have named the refresh GK12x, but fuck it! we can use that leading digit instead. So here comes GK20x.
GK104 refresh with a crazy amount of SP can easily be GK204.
Kepler has a refresh, whereas Fermi didn't really have one, it was the same chip with reworked transistors. So GF10x -> GF11x wasn't a big step. (though GF104 vs GF100 could be called a slight refresh)
Here GK104 -> GK110 is a bigger step, but it's the same basic thing with features that were omitted on GK10x.
They could have named the refresh GK12x, but fuck it! we can use that leading digit instead. So here comes GK20x.
GK104 refresh with a crazy amount of SP can easily be GK204.
I'd be tempted to easily give in and say, ok, right. But upon closer inspection, there's still too much variation - the blocks (h/v) are not only differently laid out, but have different sizes and there are some parts that are missing altogether.
What I below (1:1 c'n'p, only rotated and/or mirrored) have marked parts 1 and 2 seems to be identical between horizontal and vertical blocks but that's about it IMHO.
• vert.: there's 2 groups (1,2,3 & 4,5,6) with one block (3) being slightly enlarged.
• hor.: there's 1 group of 3 (4,5,6), one group of 2 (1,2) and one block (3) that's completely different from anything else.
[/URL]Additionally, I am missing the L2-cache altogether. If the "outer" parts of the individual SMXes are indeed 256 kiB register files each, there should be a comparatively sized portion of each MC/ROP with memory like structures which I do not see.
Well, I doubt anyone here can point exactly where the L2 SRAM cells are. But as you can see, various sorts of SRAM memory banks are scattered all over the place, partly thanks to the effects of automated layout placement. Mirrored blocks are similar (see the GF100 shot), since those are running along the same direction of the underlying metal traces and paths, but rotated features are bound to have misplaced layouts for the same reason.
iMacmatician
Regular
From CHIP Online: "Nvidia Kepler refresh: Learn to Geforce GTX 780" (original).
GK114: 2304 CCs at 1100 MHz core and 1150 MHz boost, 3 GB 6.5Gbps GDDR5 on a 384-bit bus.
It appears to be at least mainly speculation, and the specs seem quite high to me.
GK114: 2304 CCs at 1100 MHz core and 1150 MHz boost, 3 GB 6.5Gbps GDDR5 on a 384-bit bus.
It appears to be at least mainly speculation, and the specs seem quite high to me.
boxleitnerb
Regular
Sounds like nonsense. 6.5 Gbps AND a 384-bit bus? Seems like a waste somehow. Also they couldn't possibly fit that many SP at those frequencies in a 250W envelope.
And it would probably end up faster than a GK110 desktop chip at those frequencies. So I think it is a bunk rumor as well.
I wouldn't be so sure about the first part; the 2nd definitely yes.
Sure it would. Equal bandwidth, same number of ROP's running 300mhz faster, 13 SMX's at 1150mhz would have a higher throughput than 15 SMX's at 875mhz.
It doesn't matter though, rumor makes no sense. Increasing GK104 by 4 SMX units and 2 men controllers would make the chip significantly larger and I highly doubt Nvidia wants to do that. If this rumor was somehow confused with a 13 smx GK110, then it is equally weird because of the rumored clock speeds.
boxleitnerb
Regular
Could be still quite a bit smaller than GK110 though. If you sell 2 million chips and save maybe 30 bucks in manufacturing costs, you have some nice savings there. It adds up...
