nv40 300M Transistors???

[Brick_Top]

The inquirer says so

:D
http://www.theinquirer.net/?article=11002

[Deimos]

Originally posted by Brick_Top
The inquirer says so

:D
http://www.theinquirer.net/?article=11002 No they don't, they quote Economic News as saying so. That's different.

/Deimos

[MUYA]

300 Million trannies on 130nm might be a bit too much? Heat and yield issues etc. Well Uttar, Mufu or PauL can explain better :p


MUYA

[Uttar]

If we see 300M transistors on the NV40, I'll take a flight to Santa Clara and buy some expensive stuff to be able to make my voice hearable even here in Belgium and cry:
"ARE YOU GUYS TRYING TO BANKRUPT YOURSELF OR WHAT?! IDIOTS!"

I think that summarizes my opinion of that rumor pretty well ;)

300M transistors for the NV50 is not out of the question though, considering 0.09 is two times smaller than 0.13u, and they'd likely have Low-K ready for the NV50.


Uttar

[Mindriot]

wouldn't it be more likely that they would use some form of SOI (fully or partialy depleted) instead of a low-K? IBM being the fab of choice and all.

Or can they use them together?

[Paul]

Mindriot: No Low-K on NV40, and it gets less likely as the number of transistors increases.

AFAIK, NV40 is looking to have around 170-180million transistors.

I'd go as far as to say that 300million transistors is impossible at 0.13 in any mass produced part. It would be pretty much out of the question at 0.11 too.

As Uttar said, 0.09 is a possibility, but you'd probably need additional tech to make it effective re: heat and yield.

[Uttar]

NV40: SOI
R420: Low-K
NV50: SOI + Low-K

Not sure at all on that though.


Uttar

[Nutty]

How is 0.09 twice as small as 0.13 ?
Wouldn't twice as small be 0.065?

It does seem a stretch, but didn't that report say IBM only charge for working chips, not entire wafers?

[Mindriot]

Originally posted by Paul
Mindriot: No Low-K on NV40, and it gets less likely as the number of transistors increases.

AFAIK, NV40 is looking to have around 170-180million transistors.

I'd go as far as to say that 300million transistors is impossible at 0.13 in any mass produced part. It would be pretty much out of the question at 0.11 too.

As Uttar said, 0.09 is a possibility, but you'd probably need additional tech to make it effective re: heat and yield.

Now that makes much more sense.

[simwiz2]

Originally posted by Nutty
How is 0.09 twice as small as 0.13 ?
Wouldn't twice as small be 0.065?

Process sizes are line distance, but GPU cores are a square, so square each process for the core size proportions.

.13^2=.0169
.09^2=.0081

.0169/.0081=2.086

[suburbanguy]

NV30 is 125 million transistors, NV35 is 130 million transistors.


so, I'd expect the next gen NV40 to be 250~300 million.


then NV50 & NV55 (and thus XBOX2 if Nvidia provides the GPU) will probably be 500-600 million transistors.

[The Baron]

They could cool it with lasers!

......

Well, hey, c'mon, David Hasselhoff used them for seatbelts--they should be good enough for ANYTHING!

On a serious note... 300M sounds a bit iffy. ESPECIALLY after the .13u fiasco that was the NV30.

[Ninja Prime]

300m on .13 is total BS. The die would be the size of Texas! The nv35 is already large as far as chip dies go, 300 million would probably make for a die size of like 300+ square millimeters. That would not only be insanely hot, but it would be bigger than 2 P4 cores put together! Not to mention very ineffiecent use of wafer space, it would need to be like $1000 per card at least.

[Geforce4ti4200]

I doubt they can fit more than 150m on .13 however if the nv40 is .11 like rumored, they MIGHT be able to squeeze 200 million if they clock the core at like 400MHz and use serious cooling

[Paul]

Originally posted by Geforce4ti4200
I doubt they can fit more than 150m on .13 however if the nv40 is .11 like rumored, they MIGHT be able to squeeze 200 million if they clock the core at like 400MHz and use serious cooling

NV40 is 0.13, NV45 is 0.11. NV40 is 150million transistors, down from my estimate of 170million.

Thanks to Uttar for the 150million figure.

[nvnews-reader]

130million to 150 million seems like a small bump up in transistors. Can nvidia really get that much of a performance boost out of such a small increase in transistor count?

[Paul]

The key point to remember is that a lot of the groundwork for the NV40 is already there in the NV30. It really isn't a huge jump for them in terms of technology - The speed will come from them properly fixing the bugs/"features" that are holding them back in the NV3x, and increased clock speeds. More pipelines will help too, obviously.

So they don't need a GF4->GFFX transistor jump to achieve what they're going to achieve. Hell, look at the NV35 - Only 5 million more transistors, yet there is a fair amount MORE than that in terms of additions... they saved a lot of transistors by optimising paths/etc from the original NV30 design.

[Uttar]

Originally posted by Paul
So they don't need a GF4->GFFX transistor jump to achieve what they're going to achieve. Hell, look at the NV35 - Only 5 million more transistors, yet there is a fair amount MORE than that in terms of additions... they saved a lot of transistors by optimising paths/etc from the original NV30 design.

Actually, they didn't need that many transistors for the changes they did in the first place. I'll have to double check the numbers, but if some things I heard are true, a lot of things make sense: Their PS FP units are relatively cheap IMO ( unless that guy was confusing VS and PS FP units ).

Also, the amount of cache in the NV30 design is strange if it's true too. That would mean what David Kirk meant by saying "Every pathsize is optimized for 4x AA" - I think he rather meant they added 4 times more cache than what was required, lol.

I think what's important BEFORE understanding how'd they'd be able to do a huge performance increase with the NV40, is understanding why the NV30's performance is so low giving the transistor count and clock speed. We're getting there, but we ain't quite there yet.


Uttar