Latest nV info

[Joe DeFuria]

Its a tough call as to whether NVIDIA's lineup is too sparse, or ATI's is too crowded.

(Armchair mode on!)

Well, I definitely think nVidia's line-up would be too sparse (if the product line-up shown above proves to be correct.) Much like ATI's line-up had a gaping hole before the 9500 showed up.

I also think that there is "one too many" products in ATI's current line-up...and that's the 9500 non-pro. I suspect the main reason why it's in there, is because ATI wanted to quickly get something in that gap to counter GeForce4 ti sales. So I believe ATI chose the lesser of two evils...end up with an overcrowded line-up once the 9500 Pro shows up, or let nVidia continue to get sales while they wait for the 9500 Pro to ship.

I think if I were in control, when the RV350 shows up, I'd axe the 9500/9500 pro and replace it with a single product using the RV350.

Actually, I would axe the 9500 non-pro only. Keep the 9500 Pro at its current price point, ($150-$200) and release a "new" 9500 non-pro at a much cheaper MSRP. ($100).

The 9500 itself is not a bad product...It's just priced too close to the 9500 Pro right now to be viable. I can only assume they are priced so closely, because ATI can't sell it cheaper at a reasonable profit. If they can reduce the price of the 9500 to be at least $50 less than the pro, that would be sufficient differentiation, IMO. Going form 8 pipes to 4, and reducing chip size even further by going from 0.15 to 0.13, should allow them the cost savings.

So I see a good line-up for ATI being the following:

1) RV-280 @ sub $100. (The AGP 8X version of the Radeon 9000)
1) RV-350 @ $100 (Essentially the same as current 9500, only cheaper.)
2) R-300: Radeon 9500 Pro: $150-$200
3) R-300: Radeon 9700 Pro or 9700 Non-pro: $200 - $250
4) R-350: Radeon 9900: $300+
5) R-350 "Ultra": $400+ (Possible, but I doubt it.)

Notice, that I also see either the 9700 Pro, or 9700 Non-pro going away, to make room the the R-350.

 

[Dave Baumann]

The general concensus seems to be "why buy a 9500 (regular) with a 9500 pro just slightly more expensive and it performs very close to the 9700". This basically makes ATI have 2 valid products (9500 pro and 9700 pro).

Probably kinda why ATI themselves only do those two. Just let the other board vendors have a little more diversity with the tweener products.

Well, I definitely think nVidia's line-up would be too sparse (if the product line-up shown above proves to be correct.) Much like ATI's line-up had a gaping hole before the 9500 showed up.

Chip wise it seems strangly too poulated: 3 chips in one refresh, another two later on? Seems like they are going for even greater segmentation of the market with the high end part only serving the high end and not mid-end as well. If the specs are correct though, there does seem to be a hole between NV31 and NV30 though.

Actually, I would axe the 9500 non-pro only.

Why? Its pure profit. They are dumping otherwise unusable cores into this product and at least reaping some reward. It seems as though the case for dumping it on 'financial' grounds doesn't really hold much water.

 

[Bjorn]

..
2) R-300: Radeon 9500 Pro: $150-$200
3) R-300: Radeon 9700 Pro or 9700 Non-pro: $200 - $250
..

Seems like a good lineup, at least for the consumer though i'm with Russ on the replacing of the 9500 Pro.

 

[Joe DeFuria]

Why? Its pure profit. They are dumping otherwise unusable cores into this product and at least reaping some reward. It seems as though the case for dumping it on 'financial' grounds doesn't really hold much water.

That is assuming that the 9500 non-pro chips are in fact "defective" ones, otherwise unsellable. I'm not convinced that all these "failed mods" are an indicator of that. (Could be a faulty "mod process.")

But if that is the case, then yes, they may as well keep selling 9500 non-pros as long as they keep making defective chips at a reasonable clip.

It does seem very counter-intuitive to me that the higher volumeproduct would be based on a 'defective' chip.

 

[CMKRNL]

nVidia will have 4 main desktop ASIC's (in order of performance):

NV30
NV31
NV34
NV17

NV30-34 are all DX9. NV34 is missing some of the functionality of NV31, and will most likely be clocked around 250Mhz.

Competition wise, my guess is:

NV30 <-> R350
NV31 <-> RV350
NV34 <-> RV280

 

[Nagorak]

The disadvantage of that method, is that on the lower end boards, you reduce your margins significantly because you are essentially "wasting silicon". I suspect ATI is getting very poor margins on the 9500 non-pro.

How so? It's now more or less been determined that at least 3/4 of 9500 nps are junk cores that otherwise would have been ground up into sand. Depending on how you look at it, the 9500s cost ATI absolutely nothing, with their costs already being subtracted out of 9700 margins.

That is assuming that the 9500 non-pro chips are in fact "defective" ones, otherwise unsellable. I'm not convinced that all these "failed mods" are an indicator of that. (Could be a faulty "mod process.")

....................................

The failures occur with the software mod, so it can't be a mistaken soldering job. The failures are not all the same-- two cards will fail in completely different ways-- so it can't be a result of the software mod. Trust me, I have first hand experience of this hack failing.

I'm having a hard time understanding why you're trying to disbelieve this, especially since back in the day I believe you were a huge proponent of the whole "core functionality binning" theory. If you have money to waste, go ahead and buy 10 9500 NPs, I bet you get maybe 2 or 3 that are fully functional as a 9700. Even if it's as high as 50/50 (which I seriously doubt) all those otherwise useless cores are "free".

It does seem very counter-intuitive to me that the higher volumeproduct would be based on a 'defective' chip.

That's because the real product is the 9500 Pro. It costs almost the same as the 9500 but is fully functional and performs much better. Between the R9700 Pro, R9700 and R9500 Pro you have 3 lines feeding the 9500 reject line. So you end up with the 9500 being just the garbage core line, with a few supposedly fully functional chips thrown in (and who knows, maybe something really esoteric is wrong with those, or they make small rendering errors that are not visible due to being so minute).

 

[Bjorn]

It costs almost the same as the 9500 but is fully functional and performs much better. Between the R9700 Pro, R9700 and R9500 Pro you have 3 lines feeding the 9500 reject line.

Isn't it only only one line that's feeding all these products ?

But i agree, if the 9500 are bases on "junk" cores, then it's a viable product. Problem is, why don't they lower the price then to make it a viable product for the consumer ?

As Joe said, we're talking about a main stream high volume product that's supposed to be based on "junk" cores.
The yields for the 9700 must be pretty crappy for this to be a valid approach.

 

[Joe DeFuria]

How so? It's now more or less been determined that at least 3/4 of 9500 nps are junk cores...

I disagree.

In other words, I do not agree that the "modifications" being done can be assumed to be "correctly" unlocking pipelines.

IIRC, I read a few people performing the software mod on a Radeon 9500 pro, for example. And the result? Artifacts.

Depending on how you look at it, the 9500s cost ATI absolutely nothing,

That's not true either, as there is cost for just testing the chips for these specific failures...if that's what's going on.

Trust me, I have first hand experience of this hack failing.

I'm sure you have. That doesn't mean the failure is due to bad pipelines in the chip. I agree that is a definite possibility of course.

I'm having a hard time understanding why you're trying to disbelieve this, especially since back in the day I believe you were a huge proponent of the whole "core functionality binning" theory.

Actually, I was not. I was with Russ on this one...I believe that they would/could actively disable pipelines on the chip to produce another product, but I was not a proponent of testing pipelines and binning chips bsaed on pipeline failures.

That's because the real product is the 9500 Pro. It costs almost the same as the 9500 but is fully functional and performs much better.

That doesn't make sense to me.

The lower the price, the higher the volume. If 9500 non-pro is such a cheap part (with the chip being virtually for free), then why is it priced so close to the 9500 Pro?

The "real" product IMO is the 9700. Unless 256 bit bus support was an afterthought....

 

[Dave Baumann]

But i agree, if the 9500 are bases on "junk" cores, then it's a viable product. Problem is, why don't they lower the price then to make it a viable product for the consumer ?

I would have thought that would be obvious - you don't want to make it too attractive or else it will become the high volume R300 product. You want to be selling enough R300's in other products to be able to have the volume to dump bad ones into 9500 - you don't want the 9500 to be the volume product or else you'llend up wasting more good than bad on it.

And what happens to the 9500 will dempend entirely on what they decide to do with the R300 line one RV350 and R350 make an appearance - if the line of R300 product is cut back to make way for these then they may not make enough R300 cores to make the 9500, as it is, worthwhile.

 

[Joe DeFuria]

I would have thought that would be obvious - you don't want to make it too attractive or else it will become the high volume R300 product.

That doesn't make sense to me.

I've heard arguments that the 9500 non-pro is virtually "all profit", and now arguments that they don't want it to become too "high volume?"

Since when doesn't a company making such a profit on a part not want it to become high volume? Something has to give.


On a related note....

What is the "common practice" between the Fab and the customer for "bad chips"? Normally, does TSMC eat the cost for bad chips, or would ATI, or are contracts typically drawn up either way?

I mean, if this is such an obvious winfall...why does the GeForce4 MX even exist? Why not take all those 2 pipeline failure GeForce4 Ti's, and sell them insetad? Has nVidia been "paying" for failed GeForce4's, or is that TSMC's cost?

 

[RussSchultz]

They don't want it to become too high volume because they can't provide the volume (assuimng they're ressurecting dead cores).

So, you price it to (hopefully) match demand curve exactly where your inventory is completely consumed.

I'm not entirely convinced myself, since the artifacts I'm seeing don't look like its frequent enough to be a bad pipe, but I think the preponderance of the evidence is showing that the dead core idea is likely the right one.

As a small note: I still haven't seen a 9500 or 9500pro in the stores. Not at Fry's, compusa, best buy, electronics boutique, or circuit city.

 

[Bjorn]

As a small note: I still haven't seen a 9500 or 9500pro in the stores. Not at Fry's, compusa, best buy, electronics boutique, or circuit city.

As another small note, only the R9500 non pro are available in Sweden. The 64 Mb version also i might add. The 9500 Pro are supposed to arrive in 1-2 weeks but that's unconfirmed. This is online stores only i might add.

Store: www.komplett.se

 

[Joe DeFuria]

So, you price it to (hopefully) match demand curve exactly where your inventory is completely consumed.

OK, but then a disadvantage to this type of thing would be inventory management. You are trying to price it on a supply / demand curve, with a product that, I would imagine, has a very unpredictable supply.

They don't want it to become too high volume because they can't provide the volume (assuimng they're ressurecting dead cores).

And the inherent problem with that is, if they can't provide the enogh volume, they run the risk of not even being able to price it low enough to be a viable product at all (compared to the 9500 Pro.)

 

[Dave Baumann]

I've heard arguments that the 9500 non-pro is virtually "all profit", and now arguments that they don't want it to become too "high volume?"

Did you read the post Joe?

If you were putting good cores into 9500 margins would be slim - I'd even guess at non-existant. You want the 9500 to be making margins on this product they want to make as much good out of something bad, i.e. use as many 'junk' chips as possible.

IMO, this is probably why we haven't seen many 9500 PRO's either, while demand is still strong for 9700 I'd say as many R300's are going to that.

What is the "common practice" between the Fab and the customer for "bad chips"? Normally, does TSMC eat the cost for bad chips, or would ATI, or are contracts typically drawn up either way?

AFAIK, all the fab customer does is pay per wafer - they then cross their fingers and hope they yield as many good cores per wafer as they possibly can. Thats why yield is so important to the fab customer.

 

[Arun]

So you're saying nVidia is keeping the NV17? Seems strange. It would make sense to keep the NV18, but why in the world would they decide to drop the NV18? Sure, it's not better and costs more, but it's a good marketing thing. And OEMs probably love it. Or could there be inventory problems?

Also, you only give us one clock for each core. Wasn't it rumored that there was going to be two models of each chip?
My guess is that nVidia is going to make a 300Mhz NV31 & a 450Mhz NV31 ( 1200MP/s & 1800MP/s ) - but that they won't release it at the same time: A 450Mhz mainstream part is still overkill right now. They've got to wait for the proccess to mature.

As for the NV34. You say it's "missing some functionality of NV31".
Since the NV34 is supposed to also be used on the next-generation nForce, I'd guess it isn't memory-saving features such as Color Compression. And nVidia seems dedicated to support CineFX on their whole lineup.
So, I'd speculate: no adaptive texture filtering, no more shader calculators ( back to the traditional pipeline approach ) and nearly two times less temps, cache, ... ( that means FP32 is possible, but practically useless because it takes too much cache - it might even be unabled by the drivers ) and significantly slower VS ( around 60M Vertices/s, I'd guess the driver is able to determine when the VS is becoming too slow and the CPU must help )

And it would make sense to get a NV34 with even slower clocks to replace the NV17/NV18 once 0.13 matures. It might also be useful for notebooks. What about 175Mhz?
Please note that this is *speculation*


Uttar

 

[martrox]

As a small note: I still haven't seen a 9500 or 9500pro in the stores. Not at Fry's, compusa, best buy, electronics boutique, or circuit city.

Russ, bought a 9500Pro at Compusa a few months ago, and they seem to have a them most every time I go there, not a lot of them, but, then again, seems Compusa isn't carrying a lot of TI4600's either.....

 

[RussSchultz]

Guess you're lucky. I haven't seen one in its natural habitat yet.

 

[Ailuros]

NV30 <-> R350
NV31 <-> RV350
NV34 <-> RV280

I don't disagree with the guestimate as things currently look like, but you'd have to place the R300 somewhere too; unless that the R350 would replace the R300 entirely.

 

[Joe DeFuria]

Did you read the post Joe?

Yes, did you read mine? Including the sig?

If you were putting good cores into 9500 margins would be slim - I'd even guess at non-existant.

At the same time, if you were putting good or bad R-300 cores on a 9700 PCB and selling it at 9500 non pro prices, your margins would be slim, I'd even guess non-existant. Nevertheless, that's exactly what we have.

Why is ATI putting these on 9700 PCBs? Timing. They got it to market faster, despite increased cost. Can be the same reason for using "good" chips.

All I'm saying is that I certainly agree that the possibility is there for the "defective" chip theory to be true. I just don't see it as a "you must be an idiot" if you think otherwise type of situation.

IMO, this is probably why we haven't seen many 9500 PRO's either, while demand is still strong for 9700 I'd say as many R300's are going to that.

I have a different, and I think much more plausible theory.

They needed to finalize the new "128 bit bus" PCB before shipping the 9500 Pro. Why? Because apparently, they can't "disable" 128 bits of the 9700 PCB while using 128 MB of ram.

In other words, WITHOUT the new PCB, it would be physically impossible to ship the 128 Bit, 128 MB, 9500 Pro.

 

[Joe DeFuria]

NV30 <-> R350
NV31 <-> RV350
NV34 <-> RV280

I have the same question as Ailuros. Where's the R-300? It is possible that they would just discontinue it, but my guess is that from a price / performance perspective, the R-300 (in the 9700 non-pro config) will have the sweet-spot. Somewhere about $200, a step up from the NV31 / RV-350, and much less than the NV30/R-350.

I'd also like to know what happened to the NV18.

Again, the "problem" is, based on your rumored specs, it appears there might be some rather big hole in nVidia's line-up between the NV31 and the NV30. So I see this line-up:

NV30 <-> R350
??? <-> R300
NV31 <-> RV350
NV34 <-> RV280

nVidia might need a "lower clocked and cheaper" NV30 to compete directly with R-300, or alternatively "much faster than 300 Mhz" clocked NV31 to compete with R-300.