NVIDIA GF100 & Friends speculation

[fbuffer]

It's in the image center. It's not affected by the barrel distorsion because of that.

That's my point, there is distortion (where there shouldn't be any).

 

[no-X]

Maybe it's even more complex. The card isn't aligned with the image, the image isn't taken perfectly upright, the lens is very poor (it probably embodies more than just barrel distorsion).

I think because we were fooled by the puppy boards, we now see fakes everywhere despite there are none

I find more interesting the fact, that several pictures of various nVidia's product were leaked almost at the same time. I can't believe this should be just an accident.

 

[Jawed]

Hmm so what's taking up all the space?

Apart from one bit of the central area which doesn't have ROPs/MCs - and so is prolly analogue/digital outputs plus PCI Express, a lot of it repeats 4 times (i.e. once per GPC). Then there's that nice square thing occupying the centre.

In theory scaling back GF104 to GF106 should shrink size a bit more than Cypress -> Juniper - less shared logic. But maybe that's not the case... If, however, GF106 is larger because it's more than a half GF104, my bet would be an additional SM, not 256bit memory interface.

Honestly, I'm too lazy to measure stuff off the GF100 die.

I think it's possible to measure the ROPs/MCs plus the DVI/PCI-Express in the centre. Then the remainder of the central section could be scaled in one or two ways according to GPC-count. The physical stuff around the perimeter of GF100 can be scaled fairly easily too.

Then you just have to decide on the size of a GPC in GF104, versus those in GF100...

GF100 appears to have quite a bit of dead space (much like GT200 has a lot of dead space, bordering logic blocks).

Jawed

 

[Ethatron]

GF100 appears to have quite a bit of Dead Space (much like GT200 has a lot of Dead Space, bordering logic blocks).
Jawed

I knew it, it's the alien's all along!

 

[Man from Atlantis]

The memory chips are interesting. 2gbit ddr3 800Mhz 16bit. Either this board has another 4 of these chips on the back (and 2GB memory which is total overkill certainly for this performance class), or it's going to be very bandwidth constrained (64bit ddr3 interface - certainly with such a memory interface it couldn't be more than a cedar competitor no matter the die size...). Maybe the chip though would support much faster gddr5, and that's just the low-end board.

its 2gb.. 128 bit ddr3

http://www.hynix.com/datasheet/eng/...1=01&menu2=04&menu3=02&menuNo=1&m=3&s=2&RK=27

 

[mczak]

its 2gb.. 128 bit ddr3

I don't know what you're trying to tell. Sure the chips are 2gb, 800Mhz, 16bit. Bog-standard ddr3. I don't get how you infer though there are another 4 chips on the back and hence it's 128bit...

 

[DavidGraham]

Again , but this time from Tomshardware , using latest drivers , 4XAA :

1-GTX480 has a 36% advantage over HD5870 in STALKER COP @1920x1080 increasing to 58% @2500x1600 (FB limit for certain).

2-GTX480 has a a 26% advantage over HD5870 in Dirt 2 @ 1920x1080 decreasing to 13% @2500x1600 .

3-GTX480 has a 21% advantage over HD5870 in AVP @1920x1080 , but they used the benchmark , not the actual game , and @2500x1600 , the lead increases to 22% .

4-GTX480 has a 12% advantage over HD5870 in Crysis (Very High) @1920x1080 , increasing to 40% @2500x1600 (FB limit again) .

4-In COD4 GTX480 is 29% faster @1920x1080 , and 18%@2500x1600 .

http://www.tomshardware.com/reviews/geforce-gtx-460-sli-geforce-gtx-480,2694.html

And from anandtech :
GTX480 VS HD5870 :

STALKER COP : 16% @1920x1080 , increasing to 29% @2500x1600 (FB limit again)

Dirt 2 : 25% @1920x1080 , decreasing to 9% @2500x1600

BF BC2 : Only 9% @1920x1080 , decreasing to 2% @2500x1600 , however they used another test area from the game representing a worst case scenario (Water fall bench), here GTX480 came 64% faster @2500x1600 4XAA (FB again I think) .

http://www.anandtech.com/show/3836/msis-geforce-n470gtx-gtx-470-sli/5

In 4 out of 8 tests at 2500x1600 , the GTX480 enjoyed it's lead because of the frame buffer limitation in HD5870 , however that doesn't explain the advantage in 1920x1080 (maybe driver enhancements).

 

[Man from Atlantis]

http://vga.zol.com.cn/190/1903529.html

 

[Jawed]

So, only 128-bit and only 4 SMs?

Maybe it's so huge because there are two GPCs?

Why is it showing D3D10.1?

 

[Alexko]

The clocks are surprisingly high, too…

 

[Jawed]

Trying to keep up with HD5770.

 

[CRoland]

Why is it showing D3D10.1?

Maybe because only Vista SP1 is installed?

 

[mczak]

Trying to keep up with HD5770.

Apparently. And so far it looks quite promising. Ok it required high clocks (though considering what GF104 could easily do not extraordinary high, as a guess I'd suspect it will have only very slightly higher voltage), and even with that the vantage GPU score can't quite keep up with HD5770, but the same is true there for GTX460 vs HD5830... I still got some feeling though GTS450 will have trouble keeping up with HD5770 in games with that configuration but I might be wrong. Memory clock is quite modest too.

 

[Alexko]

Apparently. And so far it looks quite promising. Ok it required high clocks (though considering what GF104 could easily do not extraordinary high, as a guess I'd suspect it will have only very slightly higher voltage), and even with that the vantage GPU score can't quite keep up with HD5770, but the same is true there for GTX460 vs HD5830... I still got some feeling though GTS450 will have trouble keeping up with HD5770 in games with that configuration but I might be wrong. Memory clock is quite modest too.

Yeah, the part that really worries me is the memory bandwidth. Only 128-bit GDDR5 compared to 448-bit GDDR3 for the GTX 260, with the same number of shaders…

 

[trinibwoy]

So, only 128-bit and only 4 SMs?

Maybe it's so huge because there are two GPCs?

If the full chip is 192 shaders have we ruled out the possibility of 6 SMs, 32 shaders / 8 texture units each? I don't know how much Nvidia is playing around with the unit counts but it's a possibility and could explain the chunky die size.

 

[Jawed]

If the full chip is 192 shaders have we ruled out the possibility of 6 SMs, 32 shaders / 8 texture units each?

I was going to mention this kind of possibility, e.g. with a single despatch unit per SM that issues two instructions: 1 to MAD + 1 to either load/store or SF.

Then there's the matter of whether a GPC can support more than 4 SMs.

I don't know how much Nvidia is playing around with the unit counts but it's a possibility and could explain the chunky die size.

ATI has lower ALU:TEX for the smallest GPU, perhaps NVidia's doing the same. It seems like an amazingly heavy price to pay.

But it's worth remembering this was probably meant to be a 32nm chip. That would have made it about the same size as Juniper, I suppose. Too small for 256-bit physical interface, without doubling.

 

[Jawed]

Yeah, the part that really worries me is the memory bandwidth. Only 128-bit GDDR5 compared to 448-bit GDDR3 for the GTX 260, with the same number of shaders…

It's competing with Juniper, which has the same bus width.

 

[Alexko]

It's competing with Juniper, which has the same bus width.

Yes, but it doesn't really seem to bother it, from what I've seen it's not any more bandwidth limited than Cypress.

On the other hand, I do remember a test of the GTX 285 which showed that when you overclock either the core, the shaders or the memory by 10%, you get roughly the same speedup, meaning the card was fairly well balanced. The GTX 260 had pretty much the same flops/bandwidth ratio (I think) so this drastic reduction in bandwidth might have a significant impact.

The card shown in the screenshot only has 60.16GB/s of bandwidth, compared to 76.8GB/s for the HD 5770.

Then again it has more or less the same flops/bandwidth ratio as the GTX 460 768MB…

 

[ECH]

Alleged GTX 480-512 mini review. I looked at the power consumption and wondered if it's fact or fiction.

 

[PatrickL]

Maybe time for a GF100 POST-release thread ?