NVIDIA TSMC's 0.11 Micron Technology

[Dave Baumann]

Shocking! A company that has some services to sell has documentation bestowing the virtues of siad process? Surely not!

I think you'll also find that 130nm IC's were probably coming out of TSMC during the time that NV30 was being produced as well, however you'll likely find that they are nowhere near as large, or as complex (going by Arjan's complexivity ratio) as NV30. Thats what I mean by TSMC warning the graphics vendors that the process wasn't ready for what NVIDIA was attempting to do with it at the time.

 

[Joe DeFuria]

I don't know if TSMC has an archive section or not, but I downloaded a .pdf document from them before NV30 launched describing to customers the benefits of their 0.13 micron processes.

I'm as shocked as Dave at this.

 

[martrox]

Why do nVidia's problems in the last 2 years have to constantly be laid on the footstep of others?

It's either TSMC or Micro$oft..... yea....can't be nVidia...........

Yet ATI has had no problems with either...... why is that? Is it a conspiracy? Or did someone just make a large bunch of bad decisions.... and can't admit to it. Admission of making a mistake seems to be a fatal flaw of some companys...... and some individuals! Not to mention the inability to listen and the egotistic belief in one's own infallability.... so just who was smoking something hallucinogenic? ......

So, what's worse? Admitting to making some mistakes, taking your lumps and learning from it, or, lying, cheating, and otherwise living in total denial? Hmmmm.........

 

[elroy]

Don't forget Ati's GPU's produced at 0.13 um (RV350 and RV360) are significantly smaller transistor wise than NV30. Unless I am mistaken, R350/R360 are still on the 0.15 um process.

R300's transistor layout was hand tweaked by Ati. Fairly sure that was written in Anand's R300 preview/review.

 

[Joe DeFuria]

Don't forget Ati's GPU's produced at 0.13 um (RV350 and RV360) are significantly smaller transistor wise than NV30.

Right, though RV360 uses an "advanced" 0.13 (Low-K) as well, and RV350 IIRC was out in greater volumes / higher clockrates / lower power consumption than NV31.

Unless I am mistaken, R350/R360 are still on the 0.15 um process.

Correct.

R300's transistor layout was hand tweaked by Ati. Fairly sure that was written in Anand's R300 preview/review.

Yes, that's the prevailing thought....that this is why ATI manaaged R3xx to be similar clock rates / power consumption / transistor count at 0.15 vs. nVidia's parts at 0.13. (Which again, leads us back to a question of relative engineering or judgement competancy....)

 

[elroy]

Don't forget Ati's GPU's produced at 0.13 um (RV350 and RV360) are significantly smaller transistor wise than NV30.

Right, though RV360 uses an "advanced" 0.13 (Low-K) as well, and RV350 IIRC was out in greater volumes / higher clockrates / lower power consumption than NV31.

Agreed. But people are comparing a 120-130M transistor chip (NV30) to a 60-80M one (RV350/RV360). I'm sure those extra transistors made it difficult for nVidia to hit their target clockspeed on an unproven process. And when were we talking about NV31 (isn't that still 0.15um?)

R300's transistor layout was hand tweaked by Ati. Fairly sure that was written in Anand's R300 preview/review.

Yes, that's the prevailing thought....that this is why ATI manaaged R3xx to be similar clock rates / power consumption / transistor count at 0.15 vs. nVidia's parts at 0.13. (Which again, leads us back to a question of relative engineering or judgement competancy....)

I'd say TSMC told nVidia that the 0.13um process would be ready for a chip of that size - seems it wasn't. The question is, who's fault was it - TSMC's or nVidia's. AFAIK, no one else is producing a chip the size of NV30/35 on the 0.13um process, so it was probably a bit of both.

 

[Ailuros]

I'd say TSMC told nVidia that the 0.13um process would be ready for a chip of that size - seems it wasn't. The question is, who's fault was it - TSMC's or nVidia's. AFAIK, no one else is producing a chip the size of NV30/35 on the 0.13um process

Highlighted part: so far so good.

... so it was probably a bit of both.

Here's the point where I fail to understand the logic/connenction to the highlighted part above. It still won't explain why ATI didn't use 130nm for R3xx. Pure luck?

 

[Joe DeFuria]

I'm sure those extra transistors made it difficult for nVidia to hit their target clockspeed on an unproven process.

Oh, I'm sure it did. Too difficult in fact.

And when were we talking about NV31 (isn't that still 0.15um?)

IIRC, NV31 (FX5600) is 0.13u. NV34 (FX 5200) is 0.15u. And if memory serves, nVidia had a bit more trouble bringing NV31 to market than ATI did RV360.

I'd say TSMC told nVidia that the 0.13um process would be ready for a chip of that size - seems it wasn't. The question is, who's fault was it - TSMC's or nVidia's. AFAIK, no one else is producing a chip the size of NV30/35 on the 0.13um process, so it was probably a bit of both.

Probably right. (Although ATI is producing a chip that size on 0.15...go figure. ) I'd bet TSMC warned nVidia of the elevated risks, but didn't go as far to tell them "you're completely insane if you think you can do this." So nVidia took the risk, and failed.

 

[Ailuros]

I'd bet TSMC warned nVidia of the elevated risks, but didn't go as far to tell them "you're completely insane if you think you can do this." So nVidia took the risk, and failed.

I've got a better idea for you (conspiracy glory): ATI paid TSMC X-Million$, in order to assure NVIDIA that low-k 130nm was 100% troublefree.....since we're at it is there any position open in a conspiracy....errrr PR department at any large IHV? I just discovered a hidden talent I wasn't aware of 8)

 

[elroy]

I'd say TSMC told nVidia that the 0.13um process would be ready for a chip of that size - seems it wasn't. The question is, who's fault was it - TSMC's or nVidia's. AFAIK, no one else is producing a chip the size of NV30/35 on the 0.13um process

Highlighted part: so far so good.

... so it was probably a bit of both.

Here's the point where I fail to understand the logic/connenction to the highlighted part above. It still won't explain why ATI didn't use 130nm for R3xx. Pure luck?

Nope, I doubt it was pure luck at all. What I was trying to say was that TSMC probably conveyed to nV that the 0.13um would be ready by such a date to produce a large chip (i.e. NV30), so nV, given that they usually target the latest processes, decided to produce it at 0.13um. Obviously that decision backfired. Ati obviously decided that the 0.13um process wouldn't be ready, so stuck with 0.15um. Two different philosophies, one backfired, one didn't.

 

[elroy]

IIRC, NV31 (FX5600) is 0.13u. NV34 (FX 5200) is 0.15u. And if memory serves, nVidia had a bit more trouble bringing NV31 to market than ATI did RV360.

Thanks for clearing that up Joe. Could that be because NV31 was released before RV350? I can't remember when either was released, but I though NV31 was before RV350.

Probably right. (Although ATI is producing a chip that size on 0.15...go figure. ) I'd bet TSMC warned nVidia of the elevated risks, but didn't go as far to tell them "you're completely insane if you think you can do this." So nVidia took the risk, and failed.

I'd say so. This is the first time that a process shrink had failed for nVidia.

 

[Bouncing Zabaglione Bros.]

Nope, I doubt it was pure luck at all. What I was trying to say was that TSMC probably conveyed to nV that the 0.13um would be ready by such a date to produce a large chip (i.e. NV30), so nV, given that they usually target the latest processes, decided to produce it at 0.13um. Obviously that decision backfired. Ati obviously decided that the 0.13um process wouldn't be ready, so stuck with 0.15um. Two different philosophies, one backfired, one didn't.

I wouldn't be surpised if Nvidia insisted on using .13. IIRC at the time Nvidia only used TSMC, and had already decided that the only way to do a DX9 part was to do it at .13. We already know that Nvidia's philosophy is to rely greatly on process shrinks and memory speedups for every generation.

Nvidia were already in the mindset of relying on .13, and probably thought they would have loads of time and no competiton while they tried to perfect their design under the .13 process.

 

[Joe DeFuria]

I'd say so. This is the first time that a process shrink had failed for nVidia.

That's incorrect.

It failed them with the original TNT completely, and with the original Geforce, the parts were much lower clocked and much more power hungry than expected. If anything, I'd say nVidia has a history of having problems with new processes, and typically doesn't get it "right" until the first respin. (They also have a history of being early adopters, so it's not entirely unexpected.)

 

[Ailuros]

Nope, I doubt it was pure luck at all. What I was trying to say was that TSMC probably conveyed to nV that the 0.13um would be ready by such a date to produce a large chip (i.e. NV30), so nV, given that they usually target the latest processes, decided to produce it at 0.13um. Obviously that decision backfired. Ati obviously decided that the 0.13um process wouldn't be ready, so stuck with 0.15um. Two different philosophies, one backfired, one didn't.

Trouble is that I'm aware of how the story was, more or less as Wavey posted it; and no it wasn't just some irrelevant source exactly.

***edit:

I'd say nVidia has a history of having problems with new processes. (They also have a history of being early adopters, so it's not entirely unexpected.)

Agreed; especially with the parenthesis. It might sound "weird" yet NVIDIA has been "ironing" out manufacturing processes so far for the graphics industry. It's understandable that there's also a high risk involved.

 

[Joe DeFuria]

What I was trying to say was that TSMC probably conveyed to nV that the 0.13um would be ready by such a date to produce a large chip (i.e. NV30), so nV, given that they usually target the latest processes, decided to produce it at 0.13um. Obviously that decision backfired. Ati obviously decided that the 0.13um process wouldn't be ready, so stuck with 0.15um. Two different philosophies, one backfired, one didn't.

I don't understand. You're saying that TSMC told both ATI and nVidia that it would be ready...and nvidia took them at their word, while ATI didn't?

 

[elroy]

Nope, I doubt it was pure luck at all. What I was trying to say was that TSMC probably conveyed to nV that the 0.13um would be ready by such a date to produce a large chip (i.e. NV30), so nV, given that they usually target the latest processes, decided to produce it at 0.13um. Obviously that decision backfired. Ati obviously decided that the 0.13um process wouldn't be ready, so stuck with 0.15um. Two different philosophies, one backfired, one didn't.

I wouldn't be surpised if Nvidia insisted on using .13. IIRC at the time Nvidia only used TSMC, and had already decided that the only way to do a DX9 part was to do it at .13. We already know that Nvidia's philosophy is to rely greatly on process shrinks and memory speedups for every generation.

Nvidia were already in the mindset of relying on .13, and probably thought they would have loads of time and no competiton while they tried to perfect their design under the .13 process.

Agreed. Don't know if they thought they'd have heaps of time and no competition though. Just MHO.

 

[elroy]

I'd say so. This is the first time that a process shrink had failed for nVidia.

That's incorrect.

It failed them with the original TNT completely, and with the original Geforce, the parts were much lower clocked and much more power hungry than expected. If anything, I'd say nVidia has a history of having problems with new processes, and typically doesn't get it "right" until the first respin. (They also have a history of being early adopters, so it's not entirely unexpected.)

Agreed, but the TNT and Geforce were still competitive products. Like you said, they really shone when they were refreshed. The NV30 was not competitive with the R300 however.

 

[elroy]

Nope, I doubt it was pure luck at all. What I was trying to say was that TSMC probably conveyed to nV that the 0.13um would be ready by such a date to produce a large chip (i.e. NV30), so nV, given that they usually target the latest processes, decided to produce it at 0.13um. Obviously that decision backfired. Ati obviously decided that the 0.13um process wouldn't be ready, so stuck with 0.15um. Two different philosophies, one backfired, one didn't.

Trouble is that I'm aware of how the story was, more or less as Wavey posted it; and no it wasn't just some irrelevant source exactly.

I agree with what Wavey posted, unfortunately when I read his post I forgot to read the last line . I'll just go sit in the corner now

 

[elroy]

What I was trying to say was that TSMC probably conveyed to nV that the 0.13um would be ready by such a date to produce a large chip (i.e. NV30), so nV, given that they usually target the latest processes, decided to produce it at 0.13um. Obviously that decision backfired. Ati obviously decided that the 0.13um process wouldn't be ready, so stuck with 0.15um. Two different philosophies, one backfired, one didn't.

I don't understand. You're saying that TSMC told both ATI and nVidia that it would be ready...and nvidia took them at their word, while ATI didn't?

I should have read the WHOLE of Wavey's post before I posted, he hints that it was the other way around (TSMC told nV that it wouldn't be ready, but they targeted it anyway).

Think I need to go back into lurk mode until I can learn to read properly.......

 

[martrox]

I should have read the WHOLE of Wavey's post before I posted, he hints that it was the other way around (TSMC told nV that it wouldn't be ready, but they targeted it anyway).

Think I need to go back into lurk mode until I can learn to read properly.......

Well, elroy.......at least you are capable of seeing a mistake, admitting it & learning from it.... Wish nVidia had your keen abilities! SO..... you da MAN!