GTX/GTS460(GF104) PCB layout and spec leaked(Updated)

[mindfury]

GTX460 Specifications:
336 CUDA Core,
192bit 768MB,
Core / Shader / memory 675/1350/1800MHz.

http://translate.google.com.hk/trans...&sl=auto&tl=en






Crysis Warhead 1920x1200 VHQ 1x AA/4x AA 23.4/18.7FPS

Power Consumption about 130W

[ottoyu34]

I spot Dual 6-pin PCIE

[cirthix]

i spot nonstandard heatsink holes

[zanzabar]

that ram config looks strange, but i guess that it means that its on a 256bit buss

[ElSel10]

Man this is so confusing. I thought it was just leaked that GF104 will have 256 SPs?

[Oberon]

Rectangular packages still look odd to me.

[saaya]

lol...

first 384, then 352, then 320, now 336?
hows 336 even possible? they must have completely revamped the blocks for that as on GF100 a single block is 32 units... 336 / 32 = does not compute!
im pretty sure nvidia is seeding wrong infos again to confuse everybody and test where stuff gets leaked... hard to tell which is true and which is not...

nordichardware claims 256, but they also mention that the blocks are 16 shaders, which is wrong... its 16 blocks of 32 shaders, not the other way around...
but maybe they got this wrong cause gf104 will have 16 shaders per block...

pc inlife isnt exactly reliable, and while the pcb shot might be real, it doesnt have anything to do with the specs...
and 336 is just plain wrong... unless nvidia revamped the blocks and a block is not 16 instead of 32 shader cores...

hmmmm i believe this more than the 256 story...
336 makes sense because in the past nvidias cut down chips were always around 2/3 of the previous highend chip... 256 would be too few...
then again this is 40nm with bad yields, so nvidia might go for 256 only... but even in their geforce fx 5800U days they didnt go for a 50% cut down chip iirc...
so i think itll be somewhere around 300sps and not 256... maybe both are true, maybe the chip physically has 336sps but the 465/460 will only have 256 enabled at first, much like G92 only had 112sps at first...

[Florinmocanu]

original G92 had 112 sps Saaya :P

[zanzabar]

im just going to be happy if they come out with a smaller die, and im sure that many other people. so if people keep saying that it will be awesome of "????" happens, then every1 wants it to happen, then some1 says it will hapen it will get picked up as fact. it would make no sense for a smaller product not to be what NV is currently working on as with their thermal/power restrictions the gf100 is stupid when u could get a smaller die and clock it higher to get the same or better performance at a lower wattage and cost. and atleast with this one they got the core to shader clock right. i would expect 384 or 256 with a 256bit bus i have no idea if that would work or not but it seams reasonable


just look at the pcb, there is a ram chip out of place or in a non standard position, the pwm is short on the card, the mounting holes are really close to the power/pwm, there is alot of blank space on the IO side. but even with that we are all going to be accepting since this is what we want to happen.

original G92 had 112 sps Saaya :P

different Uarch, but it could happen

[saaya]

original G92 had 112 sps Saaya :P

my bad, typo... already fixed it, but you were faster

i wanted to add a comment about GT200...
i mean its quite interesting to compare GT200 with GT300 (gf100) since they follow the same strategy for all we know...

GT200 (65nm) was hot and had bad yields
gtx280 = 240/240 = 100%
gtx260 = 192/240 = 80%

GF100 (40nm) is hot and has bad yields
GTX480 = 480/512 = 94%
GTX470 = 384/512 = 75%

so far things dont actually look that bad at all, ~5% below what they probably aimed for, at least functional yield wise... clocks and tdp are a different thing...

but now it becomes more interesting...
with gt200 nvidia could shrink the chip to 55nm after only 6 months... which resulted in a boost in functional yields, so much that they upped the sp count of the 260 by 15%!
and upped clocks of the 280 and created a 285, PLUS it brought the tdp down enough to create a dual gpu card...

gt200b (55nm) was still kinda hot but ok and had ok-good yields
gtx295 = +2nd gpu = ~150%
gtx285 = +50mhz = 108%
gtx260 216 = 216/240 = 90%

but nvidia is still stuck with 40nm, and instead of releasing a 470 with more sps, and an updated 480 with more sps and higher clocks, they are releasing a 460/465 with LESS sps.
and if there will be a dual gpu card at all, it will probably take the place of a gtx485 they couldnt reach with a single gpu...

GF104 (40nm) probably still kinda hot but ok and ok-good yields
gtx495 = +2nd gpu = ~110%
gtx46x = 352/512 = 68%

with higher clocks the 46x will probably reach the same level as a 470... but not higher... thats why they end up calling it a 465/460 i guess, and they will probably have a dual gpu card to take the place of a "gtx485"... but other than the 295's 150% of the original gtx280, this dual card will not be notably faster than a 480 i think... at least not on average.

[Caparroz]

I spot Dual 6-pin PCIE

Two pcie connectors and 336 cores @ only 675/1350?

I spot... uhm, nevermind.

[Olivon]

Hum ... If specs are true, that's pretty really near from GTX 465 "ultra castration" card ...

[zanzabar]

its not castration if thats the new die with that as the dies specs or close to it.

[G.Foyle]

Rectangular package = weird.

[STEvil]

original G92 had 112 sps Saaya :P

my bad, typo... already fixed it, but you were faster

You're both wrong. 128.

[Olivon]

So maybe a 672 SP GTX 490 on the road ?

[annihilat0r]

8800gt had 112 shaders, the GTS 512 had 128

[ajaidev]

8800 gtx

[Florinmocanu]

You're both wrong. 128.

nope, your wrong.

8800GT with 112 sps released in october 29 2007.
8800GTS 512 with 128 sps was released in december 10 2007.

So, the first release of G92 chip was the 112 cut-down version.

[bamtan2]

that is going to be a sick card. looks like a good source

[saaya]

You're both wrong. 128.

we meant the original G92 cards, not chip... the chip had 128 but only had 112 enabled originally. i wonder if nvidia will do the same here with gf104... if 336 are enabled then it might have 384sps overall..

[xBanzai89]

If this card is priced right and beats a 5770 by a good margin I'll pick one up. Save me a lot of money messing around with this instead of a 5770 crossfire setup with having to get a new mobo. I'm hoping this a killer card from Nvidia. Price war!

[Russ_64]

GF100 (40nm) is hot and has bad yields
GTX480 = 480/512 = 94%
GTX470 = 384/512 = 75%

GTX470 has 448/512 = 87.5%

Rumored that nv will stop 470 when 465 (352sp) is released and maybe use 470 GPU's for the 490 (2x448 = 896sp) but @ 375w.........

guess they need more of a gap between the 465 and 480 to increase sales

[saaya]

doh, my bad... another typo
87.5%... interesting... same ratio as 112 to 128...

[Ket]

Looks crap. I want real results before I pass final judgement.