NVIDIA Fermi: Architecture discussion

[rjc]

Oh hotboy added an edit, as well as the frequencies(495/1100/1000) that Ailuros has above hotboy also added another sentence:

不相信的话可以翻偶以前暴的5870的分.... 现在N卡的最大问题是提升频率要靠水冷镇压...
"not believe if can overturn pairs of 5870 part...currently N card the biggest question is to upgrade the frequency depend on cold water to cool..."

Below 5870x2 performance and requires water cooling for increased frequency???

 

[Mindfury]

现在N卡的最大问题是提升频率要靠水冷镇压...
hotboy 发表于 2009-10-5 22:07

Nvdia has to use watercooling to achieve high frequency

 

[Jawed]

I'm sure Fellix will want to have a go at this: I have just scaled the die shot of GT215 so that the PCI Express blocks match in size with the GF100 die shot. This leads me to estimate 480mm² die size for GF100.

Strangely the GDDR5 interfaces on the two chips are scaling very differently, to the extent that the address/command section and the data section are scaling differently from each other. And also don't agree with the PCI Express based scaling.

Jawed

 

[Ailuros]

The online translator states "at that time" for the frequencies which if the result is fairly accurate from the translation implies that frequencies are no longer at that quite low level.

 

[MfA]

I seriously doubt it can have sufficient power consumption to make it hard to cool with a full length heatsink ... heatpipes are good at their job, for short distances (ie. across the length of a video card) water won't do much better. If you start needing a radiator much bigger then we aren't talking about 100s of Watts anymore ...

 

[trinibwoy]

Is all this still based on A1? Does that even matter for clocks?

 

[rjc]

The online translator states "at that time" for the frequencies which if the result is fairly accurate from the translation implies that frequencies are no longer at that quite low level.

The original text is: 当时 pronounced "dāng​shí​" which means usually "then" or "while" or i suppose "at that time"...but as part of the original sentence is just connecting the clocks to Vantage score previously stated.

He follows on in the next sentence that they cannot get the clocks much higher without water cooling.

Looking at original thread seems translating Lao to "old" caused a stir....probably should have used something like "Senor" or "Sir" or something a bit more gentle implying seniority somehow.

 

[Mintmaster]

I'm sure Fellix will want to have a go at this: I have just scaled the die shot of GT215 so that the PCI Express blocks match in size with the GF100 die shot. This leads me to estimate 480mm² die size for GF100.

Have you tried this methodology with RV770->RV870 or RV730->RV740? That would give you a couple reference points for 55nm->40nm.

 

[SimBy]

Oh hotboy added an edit, as well as the frequencies(495/1100/1000) that Ailuros has above hotboy also added another sentence:

Below 5870x2 performance and requires water cooling for increased frequency???

Hmm, a single overclocked HD5870 (1Ghz+ on air 100% fan speed) is already reaching close to X10000.

So if those numbers are real nV has some work to do. Not that Vantage is the best indicator for game performance, but still.

 

[Jawed]

Have you tried this methodology with RV770->RV870 or RV730->RV740? That would give you a couple reference points for 55nm->40nm.

I don't have a die shot for anything other than RV770.

Additionally, it's based on guessing that GT215 is about 11x11.9mm, something I forgot to mention earlier.

Jawed

 

[chiadog]

Translation note: 老 is more like venerable or experienced or well known sort of affectionate term rather than old which sounds kind of harsh isn't it?

Looking at original thread seems translating Lao to "old" caused a stir....probably should have used something like "Senor" or "Sir" or something a bit more gentle implying seniority somehow.

It is just an endearing term. However, lao Huang (老黄), aka old Huang, is just Chiphell's way of referring to NVidia.

 

[fellix]

I'm sure Fellix will want to have a go at this: I have just scaled the die shot of GT215 so that the PCI Express blocks match in size with the GF100 die shot. This leads me to estimate 480mm² die size for GF100.

Strangely the GDDR5 interfaces on the two chips are scaling very differently, to the extent that the address/command section and the data section are scaling differently from each other. And also don't agree with the PCI Express based scaling.

Thanks for the tip!

The data pads part of the GDDR5 interface are matching in detail and relative size between both ICs, so it's way more accurate to scale the two ASICs by those, not the PCIe interface buffers:

Almost a quarter of the size!

 

[rpg.314]

There is something fishy going on here. If gt200 is 576 mm2 is ~reticle size, then how the hell is fermi going to be ~338*3.5=~1180mm2

Also, assuming that is the rv870 die shot, I could be wrong but it looks like there are 24 simd's there.

Mmm, fish, yum

 

[fellix]

What GT200? The reddish die is GT215.

 

[jaredpace]

so 490 ish

 

[rpg.314]

What GT200? The reddish die is GT215.

What is gt215's size anyway?

 

[Scrat]

What GT200? The reddish die is GT215.

Is that a G200@40nm?

If so it would be around 250mm^2, so... what's GT215?

 

[fellix]

What is gt215's size anyway?

121,5 mm² if I remember right.

 

[Jawed]

The data pads part of the GDDR5 interface are matching in detail and relative size between both ICs, so it's way more accurate to scale the two ASICs by those, not the PCIe interface buffers:

I disagree. As I said if you compare the central part of each GDDR "64-bit" section (what I guess is the address/command bus) between the two pictures and then compare the outlying sections around those central sections, they scale differently.

121,5 mm² if I remember right.

I've never known and my guesstimation put it at about 128mm².

Anyway, roughly, GF100 appears to be in the region of 500mm² rather than 600mm².

Jawed

 

[fellix]

Looks like we have an obvious trend here: G80 (90nm), GT200b (55nm) and GF100 (40nm) -- all falling within the narrow range of ~480mm².