NVIDIA GF100 & Friends speculation

The pic you provide is exactly an GF100 SM cluster, except it has 8 TMUs instead of original 4. So why make it superscalar if the original worked without it? :)
To save transistors and die-size.

GF100 SM: 16mm2
GF104 SM: ~20mm² (Anandtech says NV increased the size by 25%)

With reduction of DP and one warp sheduler they might do such a SM even with 8 TMUs below GF100s 16mm2.
 
GT 440 (GF106) OEM :
http://www.nvidia.com/object/product-geforce-gt-440-oem-us.html
144 Cores @ 1189MHz .

And this :
001af.png
 
Looks about as accurate as this:

However , this was pretty accurate (although the timing is missed up):


96507043.jpg


Never mind the specs though (In any of these slides) , what matters IMO is the names of chips , we can make educated guesses about the rest ourselves.

There are some words in Chinese under GF110 , If anyone knows Chinese what does it say ? speculation ?
 
Here's where the image came from, and from what I seem to be getting out of the translation, they mention that a "full specification" GF110 is 768 CC, but GF110 products might have 576 CC.
I can't edit my post, but what I just mentioned is not the translation of the text under the "GF110" in the picture, it's other information from the report about GF110.
 
I highly doubt that NVIDIA can fit that much into 600mm² of silicon. I call complete BS on this one.
Even if they could pull off such a chip @ a die size <600mm² - what's the point in 786 shaders when you can only keep 720 activated due to bad yields and have to clock them <1200Mhz due to power and heat restrictions? :LOL:

Seriously, a GF104x2 card would probably be the better alternative ... I just wonder whether they'll actually try to release such a card in the wake of Cayman XT ...
 
Yes that part was mentioned already. By the looks of it, it screams "cheap". Clocks are a bit on the low side imho, so I reckon it won't be able to compete with HD5670 on the performance front, but it's only a OEM part so who cares (24 ROPs - and it's limited to 6 color pixels/clock!). At least it's got a new name, unlike the GTS450 OEM which has a very shady name imho.
And this :
001af.png
Dunno but if that's legit I think nvidia is crazy. Ok GF110 won't comment further on that. But GF112? Looks like a fully enabled GF104, with an additional ROP/MC partition it doesn't need one bit. Something similar for GF119 - a GF108 with one MC disabled. And GTX465 is placed so far above GTX460 it's not even funny.
 
Yes that part was mentioned already. By the looks of it, it screams "cheap". Clocks are a bit on the low side imho, so I reckon it won't be able to compete with HD5670 on the performance front, but it's only a OEM part so who cares (24 ROPs - and it's limited to 6 color pixels/clock!). At least it's got a new name, unlike the GTS450 OEM which has a very shady name imho.

Dunno but if that's legit I think nvidia is crazy. Ok GF110 won't comment further on that. But GF112? Looks like a fully enabled GF104, with an additional ROP/MC partition it doesn't need one bit. Something similar for GF119 - a GF108 with one MC disabled. And GTX465 is placed so far above GTX460 it's not even funny.

I don't think the height is properly representative of how powerful each chip is. GT465 looks to be that high more to do with release timings and vertical space than more power, otherwise the fabled GF110 would be much higher than the GTX480.
 
Nonsense might be an understatement. But anyhow, GF119 on 40nm does make sense in my mind, just not with so many shaders. The problem for NV is they likely don't want to implement a GF100 derivative with little architectural change on 28nm, but neither would they want to make any major change without having tested it first, potentially delaying their 28nm roadmap for reasons unrelated to the process.

So if they're clever, a last part on 40nm with a 64-bit bus (so it'd also probably be the cheapest DX11 chip from NV we'll ever get and doesn't obsolete GF108 so soon) to test architectural changes would make sense. Whether that's what they'll do, I don't know.
 
I think there is a high possibility that GF11x GPUs support more than 2 display-pipenlines (Eyefinity on Radeon).
So GF119 could be 48-64SPs 64-Bit entry part in ~70mm² with the support of 4 or more Displayports.
With two of this GPUs NV could be do a succesor of Quadro NVS cards with >=8 Displayports: http://www.nvidia.com/object/product_quadro_nvs_450_us.html

And also GF110 might have single-card support for Vision Surround.
 
And what's the point of GF119 compared to GF108? This really looks like nonsense to me.

Maybe Nvidia have a way to pack everything more densely so they are redesigning low end chips too. Also giver the low number of SPs 128bit is a very wide bus, 64bit is a better fit.
 
You also should consider that GT218s 16SPs are to little to deliver good video processing at HD solutions:
The GeForce 210 fares better than the 5450 with Smooth Video Playback here, but it still produces a rough output. Once we move up to the GT220, NVIDIA’s deinterlacer fully catches up and perfectly deinterlaces the angled lines on Cheese Slices.
http://www.anandtech.com/show/2931/4

Caicos will probably deliver enough shader power to surpass G210 at a die-size of ~70mm². GF108 with ~110mm² is to big to compete in this market.
 
The GeForce 210 fares better than the 5450 with Smooth Video Playback here, but it still produces a rough output.
Hmm, it seems, that Anandtech will never correct it. HD5450 offers the same deinterlacing quality as HD5600/5700/5800, but the pre-launch driver contained a bug, which decreased quality when "ESVP" was enabled...
 
Back
Top