R4xx will break Moore's law

[surfhurleydude]

Right....good thing the NV35 is NOT AT 500 MHZ.

I'm not saying that NV38 (assuming it comes to fruition) won't or can't be 500 Mhz+, but the fact is NV38 at 500 Mhz is indeed faster than the NV35.

Really? My old eVGA.com card said it is at 500 MHz. Hmm, I guess it was lying to me then.

Fact of the matter is, 500 MHz isn't the top of the .13 hill. If many 3rd party manufacturers are selling their NV35s at 500 MHz stock, I have no doubt in my mind that a RESPINNED core will be over 500 MHz. The NV35 core probably is capable of well over 500 MHz, but the hot 2.2 ns RAM is contributing some major heat. Which sounds logical, as when I downclocked the RAM a bit, I could then clock the core even higher than 500. Highest I got on NV35 with stock cooling was 540 core with 800 MHz RAM. If the NV38 uses some low voltage DDR II that doesn't produce as much heat as the top of the line DDR, with perhaps a tweaked cooling unit, wouldn't surprise me one bit if the NV38 is well over 500 MHz.

 

[Joe DeFuria]

Really? My old eVGA.com card said it is at 500 MHz. Hmm, I guess it was lying to me then.

Which non-overclocked eVGA card would that be?

This one: http://www.evga.com/products/pdf/n328.pdf (450 Mhz core)

or

This one: http://www.evga.com/products/pdf/N322.pdf (400 Mhz Core)

(yes, I know that eVGA actually sells some variant at 500/450...the point is, it's overclocked...)

Fact of the matter is, 500 MHz isn't the top of the .13 hill.

For the NV3x architecture....fact is, we don't know. Especially if thinkg they want to get the cooling solution down to a single slot like most normal products.

If many 3rd party manufacturers are selling their NV35s at 500 MHz stock, I have no doubt in my mind that a RESPINNED core will be over 500 MHz.

If many 3rd parties are selling overclocked NV35s stock, then I would have little doubt that they may sell overclocked NV38s over 500 Mhz too.

NV30 was RESPINNED, and the clocks went all the way from 500 to...450.

The NV35 core probably is capable of well over 500 MHz, but the hot 2.2 ns RAM is contributing some major heat.

...wouldn't surprise me one bit if the NV38 is well over 500 MHz.

It wouldn't surprise me either way. It wouldn't surprise me if the NV38 never gets released too....

 

[surfhurleydude]

Which non-overclocked eVGA card would that be?

This one: http://www.evga.com/products/pdf/n328.pdf (450 Mhz core)

or

This one: http://www.evga.com/products/pdf/N322.pdf (400 Mhz Core)

(yes, I know that eVGA actually sells some variant at 500/450...the point is, it's overclocked...)

All eVGA GeForce FX 5900 Ultras come clocked at 500/450, it's not advertised. Or at least they used to, I haven't bothered to check on eVGA's stuff in a long time.

For the NV3x architecture....fact is, we don't know. Especially if thinkg they want to get the cooling solution down to a single slot like most normal products.

nVidia's cooling designs are terribly perplexing... Every manufacturer has gotten their own single slot solution that not only cools better than the stock solution, but is much more quiet as well. Example: ASUS and MSI. They are EXTREMELY QUIET. MSI's card has a dB rating of the mid 20s. That's much MUCH quieter than the reference cooling. So, it's beyond me as to why nVidia keeps going with this double slot cooling design, especially when it has been shown that with a little bit of work, it's not needed Perhaps nVidia should employ an outside cooling designer to design a single slot cooling unit specifically for NV38?

If many 3rd parties are selling overclocked NV35s stock, then I would have little doubt that they may sell overclocked NV38s over 500 Mhz too.

NV30 was RESPINNED, and the clocks went all the way from 500 to...450.

But the question is, if many 3rd parties are selling their cards at over the reference clockspeeds, then what is the definition of overclocked? They come stock at those speeds, with a warranty of operating at those "overclocked" speeds.

You'd consider the NV35 a respin of the NV30? The thing contains more transistors, a 256-bit bus, twice the floating point operation performance, ultra shadow technology, and maybe HCT if you want to stretch things a bit, and you consider that a mere respin? R300->R350 is a respin. The R350's differences over the R300 are merely superficial. Sure it contains the shadow technology like the NV35, and it has the "f-buffer", but that's IT. Who's to say the F-buffer wasn't there in the first place and just didn't work right? The NV38 should be nothing more than a tweak and fix of the NV35 to get it to run at higher clockspeeds and possibly get better yields. It's nothing like the extreme differences between the NV35 and the NV30.

Why did the NV35 go down 50 MHz compared to the NV30? Because nVidia wanted to use a cooling unit that wasn't nearly as loud as the NV30s, and IMO is not even as loud as the Radeon 9800 Pro's (Although it's more of the pitch of the Radeon 9800 pro's thats the problem than the level of noise). Also, who's to say nVidia released it at 450 MHz on purpose so that they could later release a respin that would be more of a dramatic increase over it than if they released it at 500 MHz? That's very well a possibility.

 

[Joe DeFuria]

But the question is, if many 3rd parties are selling their cards at over the reference clockspeeds, then what is the definition of overclocked? They come stock at those speeds, with a warranty of operating at those "overclocked" speeds.

I understand your point...but who else besides eVGA is selling NV35 Ultra at greater than 450 Mhz?

My definition of "respin" is largely the same design, with a few performance and feature tweaks.

You'd consider the NV35 a respin of the NV30? The thing contains more transistors, a 256-bit bus, twice the floating point operation performance, ultra shadow technology, and maybe HCT if you want to stretch things a bit, and you consider that a mere respin?

Yes, I do. Just as I consider GeForce2 GTS a respin of the GeForce DDR, and the GeForce4 a respin of the GeForce3.

R300->R350 is a respin.

Agreed. R350 also has hardware f-buffer support, memory interface enhancements, and some stencil and FSAA optimizations.

Why did the NV35 go down 50 MHz compared to the NV30? Because nVidia wanted to use a cooling unit that wasn't nearly as loud as the NV30s,

Agreed. But shouldn't a "respin" allow them to get EITHER

1) higher clocks with the same dustbuster.
2) The same clocks without the dustbuster

Also, who's to say nVidia released it at 450 MHz on purpose so that they could later release a respin that would be more of a dramatic increase over it than if they released it at 500 MHz? That's very well a possibility.

Of course, that's a possibility.

 

[surfhurleydude]

Agreed. But shouldn't a "respin" allow them to get EITHER

1) higher clocks with the same dustbuster.
2) The same clocks without the dustbuster

Perhaps if they were the same exact thing, just respun for higher clockspeeds. The NV35 added numerous things as a trade off for the clockspeed gain. I don't see how hard that is to comprehend.

 

[phenix]

Right....good thing the NV35 is NOT AT 500 MHZ.



Fact of the matter is, 500 MHz isn't the top of the .13 hill. If many 3rd party manufacturers are selling their NV35s at 500 MHz stock, I have no doubt in my mind that a RESPINNED core will be over 500 MHz.

Maybe NV38/NV36 can go as high as 500MHz and beyond but you are not considering that next generation would probably contain significantly more transistor than these chips. More transistors would produce more heat, wouldnt they?

NV30 (125M tr)...500 MHz (dust buster)
NV35 (130M tr)...450 MHz (respin of the previous not dust buster but still not cool enough to go higher than 500MHz with sane cooling)

Consider that NV40 and R420 will use more or less comparable process technology with these and add that they may require around 150M (maybe even more) transistors which means they will be much hotter. I am assuming that they will not have infinite number of respining chance to get these chips run faster than 500 MHz, say two times at most if they are shooting for the beginning of 2004.

 

[jandar]

max speed on .13 is sounding like the same argument that we had over the speed of the R300. (275Mhz is gonna be tops for .15, just look at Matrox's Parhelia.....................)


Ati has pushed 340Mhz out of .15., what can they do with .13?

 

[Joe DeFuria]

Perhaps if they were the same exact thing, just respun for higher clockspeeds. The NV35 added numerous things as a trade off for the clockspeed gain. I don't see how hard that is to comprehend.

It's not.

How hard is it to comprehend that NV38 can be the same situation?

 

[Joe DeFuria]

Ati has pushed 340Mhz out of .15., what can they do with .13?

Correction. ATI has pushed 380 Mhz, and they're apparently not done pushing yet.

Everyone needs to understand that chips are designed with certian Mhz targets in mind from the onset. When you do little more than build upon the same basic design, there is only a certain amount of leeway in how much the clock target can change.

It's almost futile to guess "how far" ATI can push 0.13...because it depends entirely on how they choose to design the chip.

It's less futile to guess, once a chip is built and shipped...like R350 and NV35....how much of a Mhz speed bump can be gained on the same process, with the same core design.

15%-20% or so seems to be the norm per increment. I would guess 450 Mhz range for R360, and 525 Mhz range for NV38.

Once you move to a new process and / or you design a new core...all bets are off.

 

[KimB]

It's not.

How hard is it to comprehend that NV38 can be the same situation?

nVidia has yet to have a second "refresh" part that has a significant architctural change from their initial part for the generation.

Anyway, just as a point of reference, their last x8 part, the NV28, was the exact same thing as the NV25, except with AGP8x. Perhaps this will be similar...if it is an actual product.

 

[Pete]

Though, I'm still suspicious of how well the 9500/9700 thing worked out for them financially.

It must have worked out reasonably well, as nVidia has now taken up the "SE" convention. It makes sense to want to use fully functioning yet perhaps not full-speed dies, and I'm sure there's a market for such cards. Heck, you bought a 9500, didn't you?

 

[MuFu]

Consider that NV40 and R420 will use more or less comparable process technology with these and add that they may require around 150M (maybe even more) transistors which means they will be much hotter.

The use of low-k dielectrics (or perhaps SOI in IBM's case) should enable ASICs to run cooler per-clock and scale a lot higher. I've seen some estimates as high as +20-25% - not sure whether they apply to such complex parts, but a while ago I remember the UltraSPARC line benefitting quite considerable from the use of low-k materials, which must have translated to that kind of ballpark increase for a given thermal output.

Heat is still going to be a big problem though, of course.

MuFu.

 

[Megadrive1988]

We all agree the above announcement (detailing ATI's choice of Synopsis tools for development of a 200m transistor ASIC) refers to R500, right?

No.

It refers to to R420 or R4XX, not R500 or R5XX. the R420 is ment to be on .13 micron. the R500 to be on .09 micron.

ATi is doubling their chip complexity (transistor count) every generation.

R300 was around 110m transistors, so R400, or rather R420 now, should be 200m transistors or more. I would expect that R500 will be 400-500m transistors.

 

[Malo]

I would expect that R500 will be 400-500m transistors.

 

[KimB]

R300 was around 110m transistors, so R400, or rather R420 now, should be 200m transistors or more. I would expect that R500 will be 400-500m transistors.

Those numbers are growing much faster than the process technologies. What you're assuming would require that the R400 and R500 chips are going to be significantly larger in size than today's chips. I don't think that's a feasible possibility.

These counts first posted in this thread by Tim are much more likely:

0.15: 107 millions (the r350)
0.13: 142 millions
0.11: 199 millions
0.09: 297 millions

These dies would all have the same surface area (Note: this is based upon a simplification of the process technologies. ".13 micron" is just a description of one piece of the process...other pieces may evolve at different rates, meaning that these numbers are just estimates), and therefore should give decent pictures as to how high in transistors these companies will be willing to go on each process.

 

[bdmosky]

Perhaps these SE cards are PCI-Express variants??

 

[Megadrive1988]

it's obviously really, that R500 would be 400~500 million transistors.

R100 = 30m transistors
R200 = 60m transistors
R300 = 110m transistors
R420 = 200-250m transistors (likely)
R500 = 400-500m transistors (likely)


ATI, roughly, doubles its chip complexity each generation.


what process was R100 (Radeon64) on?

it was also recently reported that R4xx would be in the hundreds of millions of transistors, which suggests at least 200m.

would it be so hard to see R500 being at least 400m transistors?

 

[RussSchultz]

Though, I'm still suspicious of how well the 9500/9700 thing worked out for them financially.

It must have worked out reasonably well, as nVidia has now taken up the "SE" convention. It makes sense to want to use fully functioning yet perhaps not full-speed dies, and I'm sure there's a market for such cards. Heck, you bought a 9500, didn't you?

If me buying something was all it took for something to be financially succesful...

It wasn't a fully-functional-just-not-at-speed-die, it was ressurrecting bad dies of which half the chip was non-functional for some reason or another. neither you nor I know the details of the yield and recovery for the parts. I'm guessing the .15u yield was in the 80% range. That makes for a production ratio that is "upside down" from what is usually sold.

Maybe it would work out better for the .13, where the fully working dies yield would be in the 30% range, and you could salvage perhaps another 30-50% into saleable units.

There's lots of unknowns, and put that together with how things work at the company I work at, I'm suspicious of how well it worked out for them as a business model.

Plus, if the 9500 was so successful, why the 9600?

 

[FUDie]

Plus, if the 9500 was so successful, why the 9600?

Easy one. First, the 9600 is faster than the 9500, making it more competive. Second, to meet demand.

-FUDie

 

[RussSchultz]

Plus, if the 9500 was so successful, why the 9600?

Easy one. First, the 9600 is faster than the 9500, making it more competive. Second, to meet demand.

-FUDie

Or "cheaper to make, yet sell for the same price".

In other words, a separate die makes more economic sense than ressurrecting yield cast offs.

But again, its all speculation on our part as to whether it worked out well for them or not.