NV30 delays to 2003? [Archive]

Rookie

27-Jun-2002, 12:24

From ET (http://www.extremetech.com/article2/0,3973,286670,00.asp):

"Nvidia Corp.'s next-generation NV30 could miss this year's launch window "for all intents and purposes," a Wall Street analyst said Wednesday.

In downgrading the stock of Nvidia, Santa Clara, Calif., Prudential Securities' Hans Mosesmann said he wanted to prepare investors for the "worst-case" scenario, that shipping volumes of the NV30 or "GeForce5" could be pushed out until 2003. "

Teasy

27-Jun-2002, 12:31

Wait a moment.. was that quote really from Hans Molemann, the little freaky bloke from the Simpsons??

Gollum

27-Jun-2002, 12:45

Kyle is not the most popular guy around here lately, but he has this to add about above newsitem:

Keep in mind that all of this is coming from a stock analyst, not that I am making a call on his judgment, but I do have something to add.

We got word yesterday that that the NV30 GPU had come back from the fab on a .15 micron architecture instead of the planned .13 micron. It has been rumored that this could certainly effect NVIDIA launching a part that will be competitive with the ATi R300 which has already had almost three months to mature.

Take all that with a grain of salt, but when I heard that yesterday the first thing that went through my mind was NV30 NOT being right around the corner. Then today when I saw this it seems there may be some truth to that.

We've been assuming Nvidia would lag behind ATi but would at least have a small advantage due to a .13 micron chip, if it turns out to be .15 micron after all it will not only delay the chip, but might also make it a lot less competitive.

If this were true, this might be ATi's best chance at really undermining NVidia's role as industry leader in the long term ...

27-Jun-2002, 12:55

I agree Gollum that if this is true, its up to ATI to stand up and shine or drop the ball.

Trawler

27-Jun-2002, 13:07

If true, Nvidia has just pulled a 3dfx. :wink:

Some kind of irony to it though. Remember how 3dfx were shooting for the mature (.25?) micron setup for VSA-100, whereas Nvidia went for the riskier new process (.22?) for the GeForce? Nvidia bet right on that one, but this time they may not have been so lucky...

K.I.L.E.R

27-Jun-2002, 13:14

Regardless who's fault it is. We need to see drivers, performance and if the R300's FSAA is brocken ;)
Along with aniso quality, performance. What software is optimised for the hardware and so on.

If Nvidia release their chip 3-4 months later but their drivers are more mature than Ati's (meaning stable, faster more compat....) than that will change everything won't it?

27-Jun-2002, 13:32

Naw,

that will give the ATI fan boys 3/4 months to gloat :o

jvd

27-Jun-2002, 13:43

heh , unless of course nvidia has the bad drivers... if this is a new chip and not just yet another rehash of of the geforce they'd have to start thier drivers from almost from scratch.

Randell

27-Jun-2002, 14:07

I beleive it will be a new core.

Joe DeFuria

27-Jun-2002, 14:08

I think ATI would hold the driver advantage this time around. A DX9 part is much closer to the R-200 architecture than the NV2x architecture. So it would be fair to assume that ATI would be able to leverage their R-200 drivers more than nVidia couls leverage the NV2x ones.

I suspect both R-300 and NV30 to have their share of driver teething issues, but I think R-300 drivers will be quicker to ramp up to "par." If NV-30 is a few months later and R-300 (as is common wisdom), AND if it ends up on 0.15 micron (current speculation), nVidia would likely take clear back seat to ATI at least until the battle of the 0.13 refreshesh parts next spring.

RussSchultz

27-Jun-2002, 15:23

Wheee. the next round of speculation begins.

I'll only speculate on how long the thread will be: 5 pages.

Randell

27-Jun-2002, 15:56

hmm driver cores may bve closer to DX9 but NVidia driver team have the tarck record, and OGL is still 'theirs' isnt it?

Joe DeFuria

27-Jun-2002, 16:04

I think nVidia's driver "track record" is actually more myth than reality. Every NEW nVidia core had drivers that needed significant improvement after the release of the core.

I don't see any real difference, for example, for the way in which the R-200 drivers matured relative to the NV1x or NV2x drivers.

Now, developers have been using nVidia hardware as the "primary" development platform (deservedly so) for the past few years, so of course games tend to run with fewer "issues" on nVidia hardware. That is more of a consequence of nVidia hardware being used as the development platform, rather than the quality of their drivers.

That being said, nVidia's current drivers for the NV2x core are a bit more mature than those of ATI's drivers (especially GL). But the NV2X core is also about 6 months older than the R-200 core.

Brutal Deluxe

27-Jun-2002, 16:59

Yea people are pretty fast to say "nVidia drivers = good, ATI drivers = bad, period."

Dont forget that the GF3 got its ass kicked by the GF2U when it first came out. And the it has WAY better performance now than it did in the beginning. On top of that some of the features of the GF3 was working in the driver until september.

SteveG

27-Jun-2002, 17:12

If the nVidia driver quality is a "myth", then this myth is promulgated by none other than John Carmack:

"Nvidia's OpenGL drivers are my "gold standard", and it has been quite a while since I have had to report a problem to them, and even their brand new extensions work as documented the first time I try them. When I have a problem on an Nvidia, I assume that it is my fault. With anyone else's drivers, I assume it is their fault. This has turned out correct almost all the time."

(from his Feb 11 plan update at
http://www.bluesnews.com/cgi-bin/finger.pl?d=1&time=20020625173211 )

Randell

27-Jun-2002, 17:28

Apparently also the lack of 'pixel shaded' water in Never Winter Nights on the 8500 is due to an unfixed bug in the ATI fragment shader OGL drivers. Humus has confirmed it is not fixed.

LittlePenny

27-Jun-2002, 17:57

I think it would be interesting if NVIDIA did happen to have a two chip solution. Two chips with ~60 mil transistors could theoretically get more clockspeed than one with over 100 mil. Could be an interesting fix for TMSC's manufactoring limits.

ben6

27-Jun-2002, 18:10

This is in direct conflict with what I was told recently.... (Cg launch) and even more recently but I can't attribute the source

arjan de lumens

27-Jun-2002, 18:16

AFAIK, the main factor that limits chip sizes is yield, rather than clock speed. (with proper design, large chip size should not cause a reduction of clock speed.) Which could still make a dual-chip implementation desirable from a cost perspective.

DemoCoder

27-Jun-2002, 18:16

If the NV30 was designed for .13um, I doubt they can just recompile it for .15um and have it work. That part sounds fishy. And given the nature of the chip that we know from NVidia's mouth: 120m transistors, I doubt it is a .15u chip.

More than likely, Kyle got his stories mixed. NVidia had many new chips coming back from the fab, probably some of them are .15um (the GF5 MX?) With .13um they are probably having some bugs to fix, but I highly doubt it will run on .15um. We're talking 120m transistors, that will be hotter, slower, and a super large die size.

Of course, this is all speculation as Russ said. We will see this fall.

ben6

27-Jun-2002, 18:25

Ok this is kind of a gray area for me. On the record, Dan Vivoli said 2 weeks ago that NV30 was Nvidia's fall part. The other conversation was kind of off the record on some things but affirmed the statement that NV30 was fall part took place yesterday. Since they already commented on the record on NV30 as a fall part, I'll leave it at that.

Joe DeFuria

27-Jun-2002, 19:12

Unfortunately ben, that doesn't help much, for 2 reasons:

1) nVidia is also on the record saying the NV30 is 0.13 microns. That "on the record" statement is in fact the one in question here, so any other "on the record" statements should be treated the same way....questionable.

2) "Fall" can be a very loose definition. Dec 20th is still "the fall". ;) There's also the question of not when the product launches, but when it "ships" and in what quantity. In other words, nVidia could "launch" the part this fall (to combat ATI PR), not ship it until '03, or only "ship" a dozen boards in the fall to reviewers, while the on-line "arrival" dates get mysteriously pushed back week by week. ;)

Now, historically, the nVidia "fall" part is launched in late August / Early September, and then ships around October / November. So that's what I "hope" nVidia means by a fall part. However, most folks don't seem too confident that ANYONE could ship a 0.13 micron part in quantity (especially with NV30 complexity) in that time frame.

So, the 0.15 rumors for nV30 would not surprise me in the least. nVidia has made "back-up" chips on less advanced micon processes in the past, and shipped them. (TNT-1 is well documented, and I suspect GeForce-256 shared the same fate.) I don't believe (as DemoCoder is implying) that nVidia just took the 0.13 design and retro-fitted it to 0.15...but nVidia was probably designing both chips all along. Perhaps planned on using 0.15 only for prototyping and for such "emergencies" as we are speculating.

As for 120 million on 0.15....well...ATI is apparently doing 107 million, which most found very remarkable. What's another 13 mill or so? ;)

Joe DeFuria

27-Jun-2002, 19:23

If the nVidia driver quality is a "myth", then this myth is promulgated by none other than John Carmack:

I'm well aware of Carmack's comments....and we all know that Carmack uses nVidia parts as his reference platform. The "myth" is the internet public taking such statements in isolation and extrapolating beyond reason.

Humus also found several bugs in nVidia drivers. Point being: when you use one platform to develop you stuff on, and then port to another platform, it's not surprisng that some stuff doesn't work as expected. Humus ran into many problems when developing his code solely on Radeon hardware, then porting to nVidia hardware.

MuFu

27-Jun-2002, 20:09

So, the 0.15 rumors for nV30 would not surprise me in the least. nVidia has made "back-up" chips on less advanced micon processes in the past, and shipped them. (TNT-1 is well documented, and I suspect GeForce-256 shared the same fate.) I don't believe (as DemoCoder is implying) that nVidia just took the 0.13 design and retro-fitted it to 0.15...but nVidia was probably designing both chips all along. Perhaps planned on using 0.15 only for prototyping and for such "emergencies" as we are speculating.

Having played my part in spreading that rumour (to be honest I mentioned it quite idly and jokingly at the start), I fully agree with Joe. If they are developing NV30 for 0.15u. It will be as a back up for the 0.13u part. They know they can't count on TSMC. I think the "information" has been misunderstood by a lot of people; if it came back on 0.15u it wasn't because TSMC said "sorry no 0.13u - here's a 0.15u version to play with while you wait". It just doesn't work like that as a lot of you (especially EE people) know. :D

Fresh in my inbox today is news that a September date is still priority for a completion of NV30 R&D. If TSMC aren't ready to go by then they will "feel nVidia's wrath". I don't mean that lightly; no doubt nVidia will be extremely agressive because they really need to get NV30 shipped by the New Year. They can't allow ATi too long at the top.

MuFu.

Basic

27-Jun-2002, 21:09

We got word yesterday that that the NV30 GPU had come back from the fab on a .15 micron architecture instead of the planned .13 micron.
Oh, I can see it in front of me.

Rather ashamed EE at his boss' office:
You know, the other day I accidentaly spilled some Coke on my keyboard, and since then my '3' key sometimes generates a '5'. And I must tell this funny thing that happend when we sent the NV30 tape to TSMC.

:D

Gunhead

27-Jun-2002, 21:25

, that shipping volumes of the NV30 or "GeForce5" could be pushed out until 2003. "

"GeForce 5"? Nah nah...Something silly like "Omen" or (parheliaesquely) "Eclipse", rumour tells us.

DemoCoder

27-Jun-2002, 21:41

ushac

27-Jun-2002, 21:46

Joe DeFuria

27-Jun-2002, 22:04

It's more likely that NVidia's .15um backup is the NV28 not the NV30. They probably did both in parallel, with the NV28 on .15um being the "backup" in case the NV30 on .13um can't make fall.

The "problem" with that, IMO, is that GeForce4 is already 0.15. (In fact, the NV2x core has already been revised TWICE on 0.15. The original GeForce3 as well as the GeForce3 ti are also 0.15). I don't see the clocks going much higher for the NV2x core on 0.15.

So a 0.15u NV28 doesn't sound much like a back-up plan to me.

So I don't expect much of the NV28 in the way of bringing any new performance to the market...if it is a 0.15 micron part as you're suggesting. (The rumor is, feature wise, NV28 is just the GeForce4 plus AGP 8x support.)

I expect that nVidia could have 1 or 2 reasonable back-up plans:

1) NV30 on 0.15 - just a power hungry, lower clocked version of the planned 0.13 variant. The challenge here is to hopefully be able to clock it high enough so that it performs better than the GeForce4 Ti...while still getting decent yields and staying within power consumption limits.

2) NV2x on 0.13 - a "true" GeForce4 Ultra. (Significantly higher clocks than NV25). The 0.13 might not be ready enough to produce a 120 million transistor NV30, but it might be in good enough shape to get good yields on a 60 million transistor part.

LittlePenny

27-Jun-2002, 23:39

"GeForce 5"? Nah nah...Something silly like "Omen" or (parheliaesquely) "Eclipse", rumour tells us.
Well if they decide to go with the "Omen" then it would certainly fit with some people's claim that NVIDIA is evil.
http://us.imdb.com/Title?0075005

Eminem

28-Jun-2002, 00:38

Didn't know omens were evil. I thought they were just a sign of what is to come in the future.

BoardBonobo

28-Jun-2002, 00:54

Pete

28-Jun-2002, 01:00

Don't you know that anything nVidia associates with is evil?

psurge

28-Jun-2002, 01:37

Well I've said it before and I'll say it again - Nvidia or ATI should name their next card Bambi.

Can you imagine the reviews? "Bambi annihilates the competition"
"Bambi puts the smack down on ****". "No card on the market today can hold a candle to the awe-inspiring Bambi".

I'm serious.
Serge

RussSchultz

28-Jun-2002, 01:37

If NVIDIA can make the GF4 in .13u, there's no reason it can't make the NV30 in the same process.

Assuming your backend tools are capable of synthesising, place and routing the GF4 properly, it should be able to do the same for the NV30. Same thing with the analog side of the chip--if they can solve it for the GF4, then the NV30 should go right along. The only difference in yield would be directly related to the proportion of the die sizes. Because of this, I don't buy the "maybe a GF4 in .13u instead". We may see a GF4 .13u before the NV30, but it won't be because of the process holding the NV30 back--it will be because of something wrong with the overall design.

Also, I believe that it would not be as tough as a job to retarget from .13 to .15 as many people are claiming. It may be an optical enlargment of the mask is enough. If not that, they can certainly use different standard cell library (nearly all ASICs these days are done with standard cell) and design rules to guide their synthesis tools to retarget in less than a few days. All of their analog stuff is already designed in .15u, so there won't be any redesign effort there.

Just a two pence. :)

BoardBonobo

28-Jun-2002, 01:41

Don't you know that anything nVidia associates with is evil?

Well I guess with the amount of heat the GF4 puts out it wouldn't be out of place in hell.

'Have you ever danced with nVidia in the pale moonlight?'

We should run a poll, does god use ATi, nVidia, Matrox etc and if so why?

:D

psurge

28-Jun-2002, 01:49

DemoCoder

28-Jun-2002, 02:14

You can just "enlarge" a chip by optical zoom because the timing and electrical signal characteristics are just different as you go smaller. From what I have seen, most "die shrinks" actually require some patches and I'm sure "die enlarges" do too. Especially given the stringent timing, heat distribution, etc requirements of a 120M transistor chip.

Doomtrooper

28-Jun-2002, 03:08

Well a good example is the Geforce 4 stats that were posted here, engineering/design costs of a 160 million dollars and there was NO die shrink, it went from .15 to .15..or Geforce 3Ti to Geforce 4. For all intensive purposes the only real change in the feature set was PS 1.3 support and a second vertex shader.

Specifically on GF4 project:

63 million transistors, 78% logic

50% of RTL from GF3 modified

400k lines C , 800k lines Verilog

Schedule targets: 9 months (!!) to tapeout, 100 days from tapeout to ramp

40 to 70 engineers in implementation alone

$160 million in tools, $40 million of which on emulation

Geeforcer

28-Jun-2002, 03:10

Well a good example is the Geforce 4 stats that were posted here, engineering/design costs of a 160 million dollars and there was NO die shrink, it went from .15 to .15..or Geforce 3 to Geforce 4. For all intensive purposes the only real change in the feature set was PS 1.3 support and a second vertex shader.

Specifically on GF4 project:

63 million transistors, 78% logic

50% of RTL from GF3 modified

400k lines C , 800k lines Verilog

Schedule targets: 9 months (!!) to tapeout, 100 days from tapeout to ramp

40 to 70 engineers in implementation alone

$160 million in tools, $40 million of which on emulation

If you read the article where this came from, they estimate development costs to be around $100 million.

Doomtrooper

28-Jun-2002, 03:14

Ok only a cool $$ 100-160 million from .15 to .15 refresh.

http://www.eedesign.com/story/OEG20020612S0051

RussSchultz

28-Jun-2002, 03:34

RussSchultz

I can't say I have hard evidence to back up my feeling, but don't you think that several of the repeated elements in a GPU's design (TMU, vertex shader), would be worth doing semi/full custom? Secondly (again I'm no EE), but I find it *exceedingly* hard to believe that you can just "optically enlarge" a chip from .13 to .15 ... Because if that were possible, why couldn't you just shrink a .15/.18 to a .13? If things are that simple why is it taking AMD so long to do .13 K7's ?

Just from what I've read on the web it seems any process change requires a re-layout at least...

I think with the advent of standard cells, no fabless semiconductor company will commit to custom logic because it ties their business to a single point failure. Most companies use highly automated layout tools, also, which compresses the final layout, place, and routing to days instead of weeks.

As for the optical shrinking/enlarging: It works, but its not a panacea. Intel recently did it on their P4 processor and shrunk the die size by 10%. If you don't believe me, go do a google search and research it yourself.

http://www.chipcenter.com/eda/eda091399f.html
http://www.ebnonline.com/business/opinion/story/OEG20020503S0062 (about 4 paragraphs down)
http://www.sematech.org/public/publications/dict/o.htm

If your close enough in process process rules (0.18->0.15, or .25->.22), and you're lucky that your shrunk design won't break any rules, you can get away with it. Or you might not.

LittlePenny

28-Jun-2002, 04:05

Well I've said it before and I'll say it again - Nvidia or ATI should name their next card Bambi.

On the contrary I think naming a card "My Penis" would be brilliant. Cut from Anand's Parhelia review:In order for "My Penis" to be an attractive performer... or In the end it comes down to what sort of a value "My Penis" brings to the table. At its ESP of $399, "My Penis" doesn't deliver performance that's equivalent to what a $399 card should provide
No matter how hard (no pun) the card blew (no pun) I don't think there would be too many a reviewer on the net willing to use the cards name in a negatory remark.

Moloch

28-Jun-2002, 04:26

psurge

28-Jun-2002, 04:35

lol@LittlePenny

calling tech support would be humiliating.

Dave Baumann

28-Jun-2002, 09:04

Someone recently pointed out some of the hurdles that nVIDIA need to clear to get NV30 out on .13um; these are:

New Architecture
Largest Architecture they've done
New Process
Copper Interconnects
New dielectric processes

If this is the case then there are some issues to be solved there - I can see why ATI have removed a few of those variables for now (probably to attack those relating to .13um at a later date), however history has shown that if anyone can do this it is nVIDIA.

Most companies use highly automated layout tools, also, which compresses the final layout, place, and routing to days instead of weeks.

3Dlabs recently took us around their office to show the process that their chips go through (which no doubt differs slightly from one semi con to another depending on what EDA tools they use) but from my understanding of the process it would take some considerable time to analyse the optimal path for the layout. Also, I assume, if new dielectrics are used from .15um to .13um then that introduces a new variable which might increase the layout time.

Above

28-Jun-2002, 09:36

nVidia should maximize their return on the Geforce 4. It's only been on sale for 3 months. A lot of developers are not even programming with T&L and vertex shaders in mind, which are Geforce 2 and 3 features. No need to release the nv30 so fast. It will only encourage ATi to release their next generation part before most features have been enabled and most bugs have been dealt with. That doesn't do us any favors. Notice I use the word "most". Bah, ATi...

Randell

28-Jun-2002, 11:53

A lot of developers are not even programming with T&L and vertex shaders in mind, which are Geforce 2 and 3 features

Consumer TnL = geforce256 feature = 3 years old now.

I find it hard to believe 'a lot' of devleopers are not even programming assuming at least a Geforce 256 TnL based card. Sure software TnL paths to allow fallback to Tnt2 & Kyro (& gasp Voodoo) compatibility.

asicnewbie

30-Jun-2002, 18:51

"UMC closing in on deal to produce Xbox chips"
http://www.eetimes.com/story/OEG20011129S0040

A half year ago, TSMC's 0.15u capacity wasn't large enough to meet demand for the MS-Xbox GPU. Granted, Microsoft's projections were for millions X-Box's by the end of 2002, but even so...If all 4 GPU-makers (3DLabs, Matrox, ATI, NVidia) requested 0.13u process technology, does TSMC even *have* enough 0.13u capacity to go around?

> I think with the advent of standard cells, no fabless semiconductor
> company will commit to custom logic because it ties their business to
> a single point failure. Most companies use highly automated layout
> tools, also, which compresses the final layout, place, and routing to days
> instead of weeks

For a 'conservative' design (i.e., doesn't push die-size, clock frequency, or analog-issues), that's true. But today's 3D-GPUs are every bit as cutting-edge (manufacturing and design wise) as Intel's and AMD's best CPUs. I bet place and route on these monsters took weeks if not months of planning. (All done in parallel with other design-activities, of course.)

Nagorak

01-Jul-2002, 03:34

Didn't know omens were evil. I thought they were just a sign of what is to come in the future.

It's just another downright sterotype :evil: !

But I do think omens are signs of something nasty in the woodshed. I mean you never say "I've got a good omen about this" do you?

Actually there can be both good and bad omens. If you come across a heads up penny, it could be considered a "good omen", etc.

Omen wouldn't really make much sense for a graphics card name, however.