R4xx will break Moore's law

[phenix]

Arstechnica did a great article on Moore's Law some months back that basically covers the above with some pictures, etc. Or you could just read the darn thing yourself. I highly recommend it. To me it's perhaps the most prescient document in computer science with the possible exception of Turing's paper on AI. Of course, I'm a sucker for that sort of thing, so YMMV.

I was thinking to refer the same article. It's the best piece on the web on the issue as far as I am aware and its from Hannibal.

These two effects combine to give cost/transistor vs. transistor count a sort of U-shaped curve. The minimum of this curve is the most efficient transistor count for that particular process node. And that value is what Moore's Law predicts will double every period.

Exactly, its that the U shaped curve changes steadily by smaller process tech, larger waffer sizes, enhanced production tech more fabs etc and in 1/1.5 years you end up with the twice more maximum feasible transistor number on a single die.

For R420 I dont think it will contain more than 200M transistors hence twice as much of R300. I think (pure speculation) it was initially meant for R400 which is abondoned because it would only be feasible on better tech (e.g. 90nm, 300mm wafer). Since this technology will not be available then it was changed to R420 which will (hopefully) accomplish "twice as fast as R300" aim with fewer transistors. How they will get there is however is mystery to me.

By the way, why do you think that doubling the processing width (going from 8 pipelines to 16) would require twice as much recourses? I was thinking that a big portion of the die was used for shared components so a 16 pipelined VPU would require less than 230 million (115M X 2) transistors.

 

[Bjorn]

accomplish "twice as fast as R300" aim with fewer transistors. How they will get there is however is mystery to me.

Not that i'm especially knowledgeable in this area but i would guess you don't need to double the amount of transistors to make the card twice as fast.

Especially since "twice as fast" can be limited to specific areas. F.e with long shaders, FSAA....

 

[phenix]

accomplish "twice as fast as R300" aim with fewer transistors. How they will get there is however is mystery to me.

Not that i'm especially knowledgeable in this area but i would guess you don't need to double the amount of transistors to make the card twice as fast.

Especially since "twice as fast" can be limited to specific areas. F.e with long shaders, FSAA....

No, I didnt mean clock based fastness in terms of MHz's. I was trying to say twice as fast in processing power, which doesnot necessarily require faster transistors.
I mentioned it because I remember... I read somewhere that ATI's next generation chip will be twice as fast as R300. I know its just a PR term but in terms of processing power it can be done, even without making the chip run faster.

 

[KimB]

Well, 30nm could be available by 2005-2006, at this rate we'll be at 10nm or below by the end of the decade, that'd be like 100 atoms approx. in length... hmmm....

Yeah, at about 10nm, we're talking about single-electron transistors, due to doping restrictions.

 

[Tim]

If we assume that we Ati do not want bigger and more expensive chips the transistor budget is something like this:

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

I would be pretty surprised if R420 was made on anything better than 0.13 but twice the performance should be possible with around 140-145 transistors. one and half R350 would use 160 transistors but there is a some logic that is shared. A 12 pipeline part with around 140-145 transistors does not seem impossible and clocked at around 500 MHz it should be twice as fast as the R350. 500 MHz might seem like a lot but the RV350 could reach over 500 MHz on a relatively immature process.

On the 0.09 process we are going to see some really mind-blowing stuff.

 

[guste]

According to http://www.theinquirer.net/?article=11002 NV40 is going to be 300 million transistors? Maybe thinking that r420 will be at or around there isn't so far fetched...

"The report adds that the NV40 will have 300 million trannies and run at 600MHz, use special DDR-II memory and be produced at the IBM foundry at .13µ (micron) towards the end of the year."

Actually the NV40 thing sounds really far fetched, so I think I'll just wait until I see some actual product, and go back to watching the board quietly :? .

 

[KimB]

The Inquirer has zero journalistic integrity. I typically don't try to pay attention to anything posted at their website.

 

[guste]

Yes... yes you're absolutely correct. Still, everyone on this board reads it. It's like entertainment for geeks. You can't deny it. I mean what if the NV40 really is as they say it will be? I don't believe it and I don't think others really fall for their wild claims generally speaking, but in the back of one's mind they're thinking "what if...". They try and get people all worked up, and they do fairly well at it, it seems.

 

[KimB]

Yes... yes you're absolutely correct. Still, everyone on this board reads it. It's like entertainment for geeks. You can't deny it. I mean what if the NV40 really is as they say it will be? I don't believe it and I don't think other really fall for their wild claims generally speaking, but in the back of one's mind they're thinking "what if...". They try and get people all worked up, and they do fairly well at it, it seems.

I don't read it. It's a matter of principle. I don't like how they conduct themselves, so there's no way I'm going to give them the benefit of additional hits to justify their actions.

 

[RussSchultz]

The more I think about it, they may take the 9500/9700 to "new heights" with the RV350 core (as joe was suggesting sticking several of them together)


Then they'd have 4 products: 16x1, 12x1, 8x1, 4x1, all depending on how many of the functional groups of 4 survive. (And assuming whatever glue is required to get them functioning together isn't the yield problem).

Hit the high to low with one die. That would make their productions guys happy, plus drive the cost of their wafers down.

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

 

[Hellbinder]

aaah...

You have to add some Trannies for increased MC, and Shaders.

Which is why I said Give or take a few Million Trannies.

 

[Megadrive1988]

According to http://www.theinquirer.net/?article=11002 NV40 is going to be 300 million transistors? Maybe thinking that r420 will be at or around there isn't so far fetched...

"The report adds that the NV40 will have 300 million trannies and run at 600MHz, use special DDR-II memory and be produced at the IBM foundry at .13µ (micron) towards the end of the year."

Actually the NV40 thing sounds really far fetched, so I think I'll just wait until I see some actual product, and go back to watching the board quietly

Either way, both R420 and Nv40 will have hundreds of millions of transistors.

be it 200m+ or 300m+

 

[Dave H]

Just a quick recap of that Synopsys announcement...

MOUNTAIN VIEW, Calif., May 14, 2003 - Synopsys, Inc. (Nasdaq:SNPS), the world leader in integrated circuit (IC) design software, today announced that ATI Research, a subsidiary of ATI Technologies, Inc., has adopted Synopsys' DFT Compilerâ„¢ SoCBIST to implement the design-for-test architecture for its upcoming next-generation visual processor. SoCBIST is an extension to DFT Compiler, a key component of Synopsys' Galaxyâ„¢ Design Platform. By using SoCBIST, ATI is able to improve test quality and reduce test cost for one of their most important designs.

ATI's visual processing unit (VPU) has more than 200 million transistors of digital logic. A design of this size and complexity requires not only extremely high stuck-at fault coverage, but also thorough testing for delay-related defects, the preponderant defect type in 0.13 micron process geometries and below. Using basic scan methods, excellent delay test requires up to 6X more tester time than required for stuck-at faults, which already is at an unacceptable cost of test.

http://www.synopsys.com/news/announce/press2003/ati_socbist_pr.html

MuFu.

This may be obvious to most on this board, but just in case it's not...

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

 

[KimB]

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

Well, only thing is, it's talking about a .13 micron part. I would expect the R500 to be .09 micron (or similar).

But, at the same time, I suppose that's about the right timeframe for the very beginning of the design phase of the R500.

 

[Lezmaka]

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

Well, only thing is, it's talking about a .13 micron part. I would expect the R500 to be .09 micron (or similar).

.09 is included in the statement, it's just not singled out.

0.13 micron process geometries and below

That definately doesn't mean it'll be used for 0.13 micron only.

 

[phenix]

If we assume that we Ati do not want bigger and more expensive chips the transistor budget is something like this:

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

I would be pretty surprised if R420 was made on anything better than 0.13 but twice the performance should be possible with around 140-145 transistors. one and half R350 would use 160 transistors but there is a some logic that is shared. A 12 pipeline part with around 140-145 transistors does not seem impossible and clocked at around 500 MHz it should be twice as fast as the R350. 500 MHz might seem like a lot but the RV350 could reach over 500 MHz on a relatively immature process.

On the 0.09 process we are going to see some really mind-blowing stuff.

The way I was thinking. Only that if it is really around 145M-150M transistors it will be bigger than even NV30-NV35. Given that nvidia had severe thermal problems with running this size of a chip at 500Mhz (they downclocked back to 450MHz at NV35) I think this speed (500MHz) is more or less the maximum for this process no matter how mature it is.

And for NV40, "more than 300M transistors running at 600 MHz and using 0.13 um processes" as The Inquirer suggests seems absulutely impossible to me be it produced by IBM or not.

 

[MuFu]

And for NV40, "more than 300M transistors running at 600 MHz and using 0.13 um processes" as The Inquirer suggests seems absulutely impossible to me be it produced by IBM or not.

Yeah, that's a total crock.

MuFu.

 

[surfhurleydude]

I think this speed (500MHz) is more or less the maximum for this process no matter how mature it is.

Don't think so at all. Remember, nVidia is respinning the NV35 for the NV38, and well, I doubt they'd release it at the same clockspeeds It'll be at least 500 MHz for the core. nVidia has released cards at lower clockspeeds than their predecessors before, so it's possible that the NV40 could be the same clockspeeds as the NV35 or NV38, but I really don't think so. IBM's SOI will allow it to operate cooler and thus let nVidia get some extra speed out of the NV40 core. Remember that as of right now, the .13 process is still very primitive. No low-k or SOI have been used yet.

 

[Joe DeFuria]

Don't think so at all. Remember, nVidia is respinning the NV35 for the NV38, and well, I doubt they'd release it at the same clockspeeds

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.

 

[phenix]

What do you guys think of ATI's next generation entry level chip which will supposedly replace RV250/RV280?

I assume by the introduction of R420/RV420 ATI is planning to establish a strong ground at high end and mainstream area but how the entry level chip will be is not spoken as often.

I even havent heard any code name for it. Will it be DX9 capable, what process technology will it use, how many transistors it will be made of, when will it come etc?

This price catagory is ATI's weakest point as for today and I am sure they have some surprises to target competetors offerings.