NV36 at IBM, NV40 didn't tape-out yet.

[Arun]

Hey,

I'm a very lazy man, so I'll quote my very own GPU:RW ( http://www.notforidiots.com/GPURW.php instead:

In light of new information, I suggest everyone not to trust any of the NV36 info we had so far. A lot of it is very wrong, and that means the rest most likely is. First of all, the NV36 will be manufactured at IBM's 0.13µ process. Secondly, the NV36 will not fully replace the NV31 ( at least, not the flipchip version ). And finally, it's the NV36 which taped-out at IBM, not the NV40. The NV40 didn't tape-out yet. What market segment is the NV36 in, then? I don't know. Most likely $199-$299, with th NV31 Regular replacing the NV34 Ultra, and the Ultra being between the NV36 Regular and the NV31 Regular. No idea about that part though, it's just speculation, but those three "facts" are very reliable.

There, now I'm looking really good by assuming the NV36 was being done at TSMC


Uttar[/url]

 

[Tim]

There, now I'm looking really good by assuming the NV36 was being done at TSMC

I am for one not the least bit surprised and I am very happy that I (for once) was right. I don’t think nVidia was going to risk producing their next generation product on a process, without having any prior experience with it, again.

 

[McElvis]

Maybe the NV36 will be released as the 5900 'value' card ?

 

[ClyssaN]

hmm nv40 must tape out soon.

 

[WaltC]

I am for one not the least bit surprised and I am very happy that I (for once) was right. I don’t think nVidia was going to risk producing their next generation product on a process, without having any prior experience with it, again.

Yes, and I appreciated your pointing that out to me as well the other day in another thread... nVidia is only smart to make sure the accurate info gets "leaked" appropriately to those who have an interest in it. The last thing it needs is more rampant hype circulating...

This actually makes me feel better about what nv40 might ultimately be, as opposed to just a knee-jerk, blame-it-on-TSMC, let's-do-nv3x-again-at-IBM production cycle.

 

[Mephisto]

There we have it. Nvidia is nowhere back to their 6 month cycle at all. As Nvidia missed the June deadline for their NV40 tape-out, that part won't reach production level this year. It takes around 8 months to get a new generation to the market (NV30 taped out in august and hit the street in march, NV20 taped out in late september and hit the street in may). I fear NVIDIA took away ressources from NV40 to finish NV35 on time.

NV40 will be a spring 04 part if everything goes fine. If ATI is able to get the 0.13u R3xx refresh ready for this fall, they own the high-end again. Otherwise we will see a patt, assuming nobody takes the multichip route to bypass limited GPU power.

 

[Arun]

Mephisto: Err, IIRC, the deadline was JULY, not June.
And you're forgetting BOTH the NV20 and NV30 used a process nVidia never used before ( 0.15µ and 0.13µ, respectively ) - in the case of the NV40, as we have just learned, they'll already have used IBM's 0.13µ for the NV36 ( which explains why it took nearly 6 months from NV35 tape-out to NV36 tape-out, even though I doubt there were any real design miracles in the NV36 )

Will we get real avaibility for Christmas? I don't think so. Will we get VERY limited avaibility, like the 5900 one week ago? Maybe. Will we get review for Christmas? Unless any major problems crop up, I'd say yes.


Uttar

 

[TheMightyPuck]

This is exactly what Michael O'Hara said last week http://www.veracast.com/webcasts/twp/growth-forum-2003/79405386.cfm

 

[WaltC]

Will we get real avaibility for Christmas? I don't think so. Will we get VERY limited avaibility, like the 5900 one week ago? Maybe. Will we get review for Christmas? Unless any major problems crop up, I'd say yes.


Uttar

I think this all depends on what nv40 actually is. If it's the "clean" design I'd like to see that trashcans most of nV3x, I'd say spring '04 is a likely target. If it's nv30/5 with some minor tweaks which might possibly ramp a little higher in MHz, then I'd say your guess is as good as anybody's.

As well, I think things will depend on if and when nv35 can ramp up in yields. If by August the picture hasn't materially improved then I think the chances are better for nv40 at least making an appearance by Christmas, even if it's not shipping. The real question for me is "What is nv40?"

 

[Arun]

Okay, so...

NV40:
- 550-600Mhz core clock on IBM's 0.13µ process
- 700-800Mhz GDDR2 on a 256-bit memory bus
- 8 *true* pipelines, no possibility to get more "zixels" as on the NV30 ( no idea about what's in each pipeline )
- Unified ISA/Constructs for the VS/PS and shaders 3.0. compliant
- No more FP16 / Partial Precision ( *full* FP32 ) -> likely little or no register usage performance hit
- Includes a PPP ( Programmable Primitive Processor )

I'd bet it'd feature an industry-leading pricetag, too... ( read: VERY expensive! )

Of course, that's just rumors. It could be all wrong, for all I know, hehe. But my point is that this is what I believe the NV40 is likely to be right now.
I'm sure ATI isn't sleeping either...


Uttar

 

[Ailuros]

700-800Mhz GDDR2

I assume that's a typo; considering that it should have read according to your own expectations 3 instead of 2.

no possibility to get more "zixels" as on the NV30

Hmmm...is that a "fact" so to say or just an estimate? Assuming I accept the unified shader grid, will f.e. PS be completely de-coupled from stencil fillrates?

Of course, that's just rumors.

To be honest I have some personal doubts what the unified shader grid concerns (albeit from my understanding it'll make a lot of sense) and the PPP, but I am in no way in a position to know for sure either...

 

[Arun]

Ailuros:
1. I wasn't sure if it is GDDR2 or GDDR3. Now I'm pretty sure it's GDDR2.

2. It's a rumor - so someone considers it as a fact. It makes sense to me considering the useless complexity of the NV3x's pipeline combining approach, but I don't know for sure.

3. Oopsy, sorry. Wasn't clear enough I fear. What I mean by "Unified Shading Model" is *not* dynamic allocation. I'm pretty sure that's not in the NV40 now. What I'm talking about is the fact that the VS and PS execution units are identical ( or very near identical ) and that the instruction sets were identical ( as well as *maybe* the same instruction lenghts limits, but I'm not sure about that )

4. I'd be very surprised if there wasn't a PPP in the NV40. It was one of the original targets for the NV30 I think ( hard to proof, though ) and

As I said, it's all rumors Not like I'm teaching you anything there, hehe.


Uttar

 

[horvendile]

There we have it. Nvidia is nowhere back to their 6 month cycle at all.

I thought consensus was that products would come farther apart from here on. I might have got that wrong though; it may be that generations will come farther apart but refreshes will continue to arrive with 6 months intervals.

 

[Sabastian]

There we have it. Nvidia is nowhere back to their 6 month cycle at all.

I thought consensus was that products would come farther apart from here on. I might have got that wrong though; it may be that generations will come farther apart but refreshes will continue to arrive with 6 months intervals.

I have always thought that nvidia was on a nine-twelve months product cycle. The way I look at this is that a variation of a core is not really a new product. So in effect nvidia has produced only 3 new cores with varieties of refreshes inbetween new core launches. There is good reason to argue that indeed the augmentations are new products, but the real r&d investment goes into the main core. I see the entire Geforce 1-2 lineup based off the original geforce 256 core. NV20-30 is based off the geforce 3 xbox GPU and the NV30. It was 2000 when nvidia released the original nv10 core and now it is 2003 and we have the nv30. The Geforce 3 really was slightly ahead of its time though likely because of the pressure that MS put on Nvidia to get xbox hardware asap. The same with ATi and their R100, R200 and R300 cores. The R100 was a response to the NV10, R200 the NV20 and the NV30 a response to ATi’s R300. Anyhow that is my take on things some may have bones with it though.

 

[Josiah]

what's the point of having a PPP if you don't have dynamic allocation?

 

[Arun]

what's the point of having a PPP if you don't have dynamic allocation?

Well, according to the exact same logic, I could ask you why current GPUs got a Vertex Shader since they don't have dynamic allocation


Uttar

 

[Sabastian]

I forgot to mention the reasoning behind that post. By the logic of nvidia introducing a new core on a 9-12 month bases the NV40 ought to tape out in (assuming the nv30 was taped out in Nov 2002) the Aug-Nov range this year. That is my estimate. I would be biased for an earlier tape out as a result of how poorly the NV3X has been received. My best guess.... Sept on the .13um process with Low K. unless of course nvidia was looking for a smaller process for the NV40..... then it might be a little longer. If the NV40 has already taped out or they are trying to tape it out immediately it would be an indication that they have given up on their NV3X architecture at a horrible expense.

 

[Arun]

1. The NV30 taped-out in August 2002
2. IBM doesn't have Low-K ready yet
3. Considering that each level ( high-end, mid-end, ... ) of the NV3x will only be in stores for 1 year or less, I'd say nVidia has completely given up on that architecture for a long time already, and are just in damage control mode ( in fact, being able to beat ATI in some areas with the NV35 is already an accomplishement, considering how bad the NV3x architecture is )


Uttar

 

[Sabastian]

1. The NV30 taped-out in August 2002

Well it sure did take them a lot longer then the usual time to bring the architecture to the market then.(IIRC 90-120 days.) I assumed that it was Nov when they had the first working silicon of the NV30 as we didn't see the chip shipping until ... Feb? If the NV30 taped out in Aug it took them twice as long as usual to bring to the market. I was suspicious of the NV30 in Nov @ computex, it wasn't working well then and speculated that possibly it indeed had just taped out. I don't really dispute that it was taped out in Aug though and if that is the case then indeed they could very well be attempting to tape out soon.

2. IBM doesn't have Low-K ready yet

Yeah that is something I was thinking about as well, they might very well have designed the NV40 with the .13um low k process in mind. Wouldn’t that hold up a tape out?

3. Considering that each level ( high-end, mid-end, ... ) of the NV3x will only be in stores for 1 year or less, I'd say nVidia has completely given up on that architecture for a long time already, and are just in damage control mode ( in fact, being able to beat ATI in some areas with the NV35 is already an accomplishement, considering how bad the NV3x architecture is )

Uttar

Yeah, the NV35 was released only 2-3 months after the NV30 shipped, I pretty much agree with #3.

 

[Arun]

2. IBM doesn't have Low-K ready yet

Yeah that is something I was thinking about as well, they might very well have designed the NV40 with the .13um low k process in mind. Wouldn’t that hold up a tape out?

I doubt that, but who knows.
nVidia isn't behind schedule for tape-out yet anyway, don't know where that idea came from


Uttar