iMacmatician
Regular
NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
I can't see how it could be anything else than a rebrand.
The 610M looks like a straight rebrand of the 520MX, though I think it is possible it is also a replacement for the 520M (as the clocks are "up to" the 520MX reference level).
In the exact same way the 630M looks like a rebrand of the 540M but could also replace 525M (the 550M, 540M and 525M designations were stupid anyway with only a 10% difference between each of them - interestingly the 630M does not cover the highest clocked 550M part).
The 635M is either a rebranded 550M (the 96 cores "up to" 753Mhz version) or 555M, though the 555M could already have been anything with its "up to 144 cores" and otherwise wildly varying specs (clocks, memory interface width, memory type...). Interestingly, despite claiming support for gddr5 the max bandwidth specified has gone down compared to 555M, seems to be for 192bit ddr3, or really low clocked 128bit gddr5.
So I guess the parts got shifted around a bit, and 3 models possibly only now instead of 6 for lower end, but apart from new stickers there seems to be nothing new (that is still all old 6 versions are easily possible). Or it is possible the lower clocked parts are really gone in the 600 series (which might make sense too at least for the low end 520M part which probably isn't any faster than HD4000).
GF117: 28nm and PCIe Gen3?
Could it be so small, that Nvidia will replace the ~80-90mm² GF119 GeForce 610M with it?
[strike]
btw.
107x is a Quadro device-ID.
http://forums.laptopvideo2go.com/topic/29057-v28586-windows-7vista-64bit-nvidia-mobile/
Could it be so small, that Nvidia will replace the ~80-90mm² GF119 GeForce 610M with it?
[strike]
btw.
28nm shrinks of GF119? Maybe 32-Bit GDDR5 to reduce pad-limit?[/strike]
107x is a Quadro device-ID.
Last edited by a moderator:
I remember Charlie mentioned a 28nm tiny GPU named GK?117, whiched lacks display controller. It makes sense. I think Nvidia can use such GPU together with Atoms or cheap mobile Intel CPUs (something like ION3). Such GPU doesn't need a display controller, because Atom already has one. Lack of output wiring can also reduce pad limitation. Just my speculation 
According to this table GF117 could be also replace the recently launched GeForce GT 630M:
http://www.computerbase.de/news/2011-11/details-zu-anstehenden-28-nm-grafikloesungen-von-nvidia/
There could be also some rebranding in GeForce 600M high-end:
http://www.computerbase.de/news/2011-11/details-zu-anstehenden-28-nm-grafikloesungen-von-nvidia/
144SPs and 64-bit GDDR5 @ <100mm² could be possible for GF117. GK107 seems to be a 128-bit ~150mm² part, which scales to GTX 550 Ti performance level.
There could be also some rebranding in GeForce 600M high-end:
http://forums.laptopvideo2go.com/topic/29057-v28586-windows-7vista-64bit-nvidia-mobile/
iMacmatician
Regular
Someone on these forums also mentioned an (IIRC) Optimus-only 28 nm Fermi that is ≤ GF116 specs. If the GK117 was actually a GF117, then it would make more sense name-wise than GK117 first then GK107 second.
Seems sensible enough spec-wise.
And then route display output data back through the PCIe 1x?
That's what Optimus already does.
There could be also some rebranding in GeForce 600M high-end:
http://forums.laptopvideo2go.com/topic/29057-v28586-windows-7vista-64bit-nvidia-mobile/
http://www.chiphell.com/thread-369427-1-1.html
GTX 675M = GTX 580M
GTX 670M = GTX 570M +25MHz base-clock
And GTX 560 SE as stock cleaner for GF114: http://www.qk123.com/2012/02/26/review-nvidia-geforce-gtx-560-se-part1/
Works since nearly a year, too - GTX 460 v2.
Looks like there are some issues with the 192-bit 1GB trick:
and for the 99th percentile frame time:
http://techreport.com/articles.x/22473
The frame drops on the 550 Ti probably made it less "playable" overall than GTS 450 that has only 128-bit bus and 16 ROPs, plus lower clocks.
Could it be that the 550 Ti had some inherent problems with the 192-bit 24 ROP/1GB trick, in that it performed worse than GTS 450 in these critical areas? Is it a problem that Nvidia is well aware of, but hoping that the trick magically extends to a high frame rate count right under our noses?
It is possible. I think anand tried to figure out what effect this 192bit/1GB arrangement had but with cuda you could only use 768MB so exact benchmarks weren't even possible.
The easiest way to support such a 1GB card would be if the memory was simply configured as 768MB/192bit + 256MB/64bit. As long as not more than 768MB is used that's great but after that you'd better put only things in the upper 256MB which don't require too much bandwidth.
GF117?
http://www.geforce.com/hardware/notebook-gpus/geforce-gt-620m/product-images
Anandtech thinks it is a 28nm GF108:
http://www.anandtech.com/show/5697/...es-keplers-and-fermis-and-die-shrinks-oh-my/2
.. but. NV lists a new 144SPs 525MHz GT 630: http://www.geforce.com/hardware/notebook-gpus/geforce-gt-630m/specifications
http://www.geforce.com/hardware/notebook-gpus/geforce-gt-620m/product-images
Anandtech thinks it is a 28nm GF108:
http://www.anandtech.com/show/5697/...es-keplers-and-fermis-and-die-shrinks-oh-my/2
.. but. NV lists a new 144SPs 525MHz GT 630: http://www.geforce.com/hardware/notebook-gpus/geforce-gt-630m/specifications
iMacmatician
Regular
I got the pictures of the GT 540M (GF108) and the GT 520M (GF119) and the listed die sizes on Wikipedia (79 mm^2 and 116 mm^2). Then I compared the package sizes (with the assumption that the green block around each package is the same size for both chips) with the listed die sizes using a linear regression, and used it to estimate the die sizes of the GT 620M (GF117?) and GTX 660M (GK107) from their pictures.
I got GT 620M die size ≈ 106 mm^2 and GTX 660M die size ≈ 121 mm^2. If these numbers are anywhere near close (the GK107 number doesn't seem too far off the rumored numbers at least), then the "GF117" seems too large to be a straight GF108 shrink. Either it has more than 96 CCs or they added more ROPs (4 » 16) to put it in line with GF104/114 and GF106/116, although I don't know how much space those take up.
I got GT 620M die size ≈ 106 mm^2 and GTX 660M die size ≈ 121 mm^2. If these numbers are anywhere near close (the GK107 number doesn't seem too far off the rumored numbers at least), then the "GF117" seems too large to be a straight GF108 shrink. Either it has more than 96 CCs or they added more ROPs (4 » 16) to put it in line with GF104/114 and GF106/116, although I don't know how much space those take up.
If you are looking on the picture of the backside and the solder points, GF117s package seem much smaller.
I scaled it on base of the package markings (the white cross and the white square) and the electrical components on the package.
On base of this GF117 looks to be around 60mm², which fits for a 28nm GF108 without display connectors. Although it seems a bit small for 128-bit DDR3.
iMacmatician
Regular
Good catch. I made another estimate by lining up the electrical "dots" and got ≈63 mm^2.
Although it seems a bit small for 128-bit DDR3.
Without monitor connectors, there is plenty of edge to fit 128-bit DDR3 in there.
I suppose that's the point. The smallest possible chip that can fit a 128-bit interface to really cheap ram. (And DDR3 is *really* cheap these days.)
Also explains the PCIe3 -- needs the bandwidth to move drawn frames to main memory, especially if they have had to make the PCIe bus more narrow to fit the memory interface.
Similar threads
- Deleted member 2197
- Architecture and Products
