NVIDIA GF100 & Friends speculation

[ferro]

And why would nVidia wait for a response of AMD?
Do you forget what happend with GT200? It would be really stupid to ignore the fact that amd could and i think will reduce the price of their cards. A GTX470 for 399$ with the performance of 5870 looks not interessting after a nearly 100$ price cut of the 5870...

That's why I think they will need something with significant better performance (and features) than a 5870. My assumption is that they have been targeting for this. This way they can price it higher, and everything will be fine. Neither company will benefit from a new price war. Especially when availability for 40nm parts in general is low. Healthy margins are required.

 

[Psycho]

http://electronika.vacau.com/index.php/pc-hardware/6-gpu/8-fermi?start=1
Any truth to these benches?

No, he has to learn to fake first
Putting some idea of performance gain into a spreadsheet without jittering your results is not a good idea. Relative performance of gtx480 from his table (same row/columns):

1,29	1,61	1,25	1,28	0,75
1,22	1,17	1,16	1,15	1,21
1,50	1,50	1,50	1,50	1,50
1,67	2,12	1,88	1,67	1,17
1,95	2,25	2,11	1,59	1,14
2,33	2,08	2,16	1,88	1,41

So it's always exactly 50% faster than gtx295

 

[Sontin]

That's why I think they will need something with significant better performance (and features) than a 5870. My assumption is that they have been targeting for this. This way they can price it higher, and everything will be fine. Neither company will benefit from a new price war. Especially when availability for 40nm parts in general is low. Healthy margins are required.

nVidia should sells all GF100 cards at a higher price point... Is this a new style of competition?
I guess you mean something like this:
5870 - 399$
GTX470 - 429$

But why do you think that amd would not cut their price?

 

[compres]

And nVidia was the winner of the last war.

Seriously, what are you trying to achieve here? You could be more persuasive if you showed some kind of balance in your arguments.

Are you just trolling? I bet many users here have you on ignore already...

 

[Sontin]

Seriously, what are you trying to achieve here? You could be more persuasive if you showed some kind of balance in your arguments.

I guess you should check the facts. The "first war" began in november 2007 and ended in september 2009. In this time "ATi" has not gain marketshare in the diskrete desktop segment, lost money because of selling their notebook cards at a very low price (look at q2 2009) to earn marketshare and don't make any real profit.

Are you just trolling? I bet many users here have you on ignore already...

Maybe, but i'm not the guy who make statements like "they can't win this war".

 

[Jawed]

Maybe a 17th is for compensating die-defects.

The concensus is precisely that There is a patent document for this, too.

Too bad I didn't link those really good sources, I think it was Dirk talking about how they were going through all blocks and removing redundant logic; wild shot: ~15% it was I think.

http://www.rage3d.com/image.php?pic=/reviews/video/atirv770/architecture/pics/hartog/4.jpg

Interesting, the DOT-logic is reduced to another FMA there.

DOT is crossing lanes as well as using MAD. (In ATI's current architecture there's FMA distinct from MAD - DOT based instructions are all using MAD.)

I wonder if the DOT-logic can be split into independent MUL,MUL,ADD operations. If (I don't know) x,y,z,w all can do a DOT in 1 clock, there is way more FLOP-logic in the chip available than utilizable.

You do realise there are 5 MADs running in parallel there, don't you? With adders between pairs X,Y and W,Z and between X,Z. That's how the hardware can do a DOT4 per clock.

I can't remember the page I saw the schematics, so I just made a custom one. I haven't made one for (14) years, so be humble.

I don't feel like doing a drawing:

   A  B  C  D  E  F  G  H
   |  |  |  |  |  |  |  |
   *__|  *__|  *__|  *__|
   |     |     |     |
   |  0  |  0  |  0  |  0
   |  |  |  |  |  |  |  |
   X  |  Y  |  Z  |  W  |
   |  |  |  |  |  |  |  |
   +__|  +__|  +__|  +__|
   |     |     |     |
   +_____|     +_____|
   |           |
   +___________|
   |
 DOT4

Uh, so how do you know an "exception" occured in the first place?

Dumb test after the instruction. Won't slow you down too much if there's only certain instructions you're expecting to encounter woolly stuff.

Jawed

 

[ferro]

nVidia should sells all GF100 cards at a higher price point... Is this a new style of competition?

No, this is not new. If your product is better, you can price it higher than the competition.

 

[ferro]

Maybe, but i'm not the guy who make statements like "they can't win this war".

I think you are referring to me here? I said that Nvidia cannot win a price war. My reasoning was:

1. An RV870 chip is smaller than a GF100 chip, and therefore cheaper.
2. An RV870 board is less complex than a GF100 board, and therefore cheaper.
3. An RV870 board has less memory than a GF100 board, and is therefore cheaper.
4. RV870 has been in volume production longer than GF100, and is therefore cheaper.

Therefore I believe that Nvidia should not engage in a price war, but should differentiate in performance and features.

 

[NathansFortune]

And why would nVidia wait for a response of AMD?
Do you forget what happend with GT200? It would be really stupid to ignore the fact that amd could and i think will reduce the price of their cards. A GTX470 for 399$ with the performance of 5870 looks not interessting after a nearly 100$ price cut of the 5870...

That's a lot of assumption. The price of GTX470 isn't set and is likely going to be higher than that, and the performance is not known either so to say that Nvidia's salvage part is going to have similar performance to ATi's top-end single chip is a massive assumption.

I'm not saying it won't happen like that, but the evidence points more in the direction of high-price medium performance right now. We have to see which way it goes in a few days when benchmarks start to leak. I hope I'm wrong, really I do as ATi will have to shape up with their pricing and then I'll be able to afford a decent replacement for my 8800GTX.

 

[Sontin]

No, this is not new. If your product is better, you can price it higher than the competition.

And what happen to "outdated" products? They will put in a lower market segment.

I think you are referring to me here? I said that Nvidia cannot win a price war. My reasoning was:

1. An RV870 chip is smaller than a GF100 chip, and therefore cheaper.
2. An RV870 board is less complex than a GF100 board, and therefore cheaper.
3. An RV870 board has less memory than a GF100 board, and is therefore cheaper.
4. RV870 has been in volume production longer than GF100, and is therefore cheaper.

I know this. But to win the price war also means to sell (a lot) more units than your competition.

Therefore I believe that Nvidia should not engage in a price war, but should differentiate in performance and features.

nVidia has no control over AMD. Do you think nVidia wanted the price war in july 2008?
Don't forget AMD in your assumption.

That's a lot of assumption. The price of GTX470 isn't set and is likely going to be higher than that, and the performance is not known either so to say that Nvidia's salvage part is going to have similar performance to ATi's top-end single chip is a massive assumption.

Really?
8800GTS - HD2900XT
3870 - 8800GT (the 8800 was 20% faster than the 3870...)
4870 - GTX260

I think: No.

 

[rpg.314]

But to win the price war also means to sell (much) more units than your competition.

I like this bit.

 

[Florin]

I know this. But to win the price war also means to sell (a lot) more units than your competition.

Ironically, this can also be a great way to lose the price war.

 

[Razor1]

depends on how many units are sold and the profit per card gained.

 

[ferro]

I know this. But to win the price war also means to sell (a lot) more units than your competition.

That's not what I meant though. I was just talking about prices, not about sales. AMD set the starting prices with the HD58XX, Nvidia comes with prices that are lower than AMD, forcing AMD to lower its prices, and so on. AMD can always win this game because their products are cheaper to produce.

nVidia has no control over AMD. Do you think nVidia wanted the price war in july 2008?
Don't forget AMD in your assumption.

My first "price war" response was to your comment about not taking the rumored $299 price for the GTX470 too seriously. I agreed with you, for the reason that I believe that Nvidia should not start a price war. If AMD starts one after the GTX470 release, Nvidia may have no choice, but I believe they should not start it.

 

[Florin]

AMD can always win this game because their products are cheaper to produce.

The first part isn't true but the second part is.

 

[ferro]

The first part isn't true but the second part is.

Care to elaborate?

 

[Ethatron]

The SSE unit is not shared in the sense that each core had it's private 128 bit ALU in barcelona. In BD, the two ALU's from neighboring cores are taken and put together in center of the module. So while the fp instruction stream from both cores is muxed into a single 256 bit ALU, it is not sharing in the sense of 2 cores having to do with less since the per core FP throughput is the same.

Yes, exactly. Troughput is the same per module. But can be half per thread. But this "half" is already as wide as on K10 (namely 128bit). That means per-thread throughput is the same as K10 in the worst case and may be twice K10 in the best case.

Our differences are over the semantics of "sharing" I guess.

I didn't want to imply there is a shared FPU on K8/K10. The FPU widened from K8 64bit to K10 128bit. But both dedicated.

This is not the same case in BD. The FPU technically didn't widen to 256bit as an unsplittable dedicated piece. I have the high hopes the BD design is truly designed future-proof giving the ability to make additional variations possible (including bogus ones):

- 4x int cores, 2x 128bit fpu cores splittable into 4x 64bit SIMD blocks
- 4x int cores, 2x 256bit fpu cores mergable into 1x 512bit SIMD block, or 4x 128bit SIMD blocks

And so on. That would truly be modular, and would give rise to custom customer-designed BD in the spirit of the PowerPCs from Motorola (sorry FreeScale ). A mobile BD with 8:1 INT:FPU relation for example. Why not.

But anyway, this is a GF100 thread, I don't want to highjack.

 

[Razor1]

Care to elaborate?

it depends on how many cards they sell for the amount of profit per card

50 sold for $10 profit gives you $500
25 sold at $20 profit gives you $500

but if at $10 profit you sell only 30 units, in that case you just lost money. You should have left it at $20 profit.

This is simplistic because you still have to look a market saturation of the product, competitive products, consumer "spending" levels for a given product, etc, etc.

 

[Florin]

Care to elaborate?

Nvidia has no debt.

 

[Ethatron]

DOT is crossing lanes as well as using MAD. (In ATI's current architecture there's FMA distinct from MAD - DOT based instructions are all using MAD.)

Ah, true.

You do realise there are 5 MADs running in parallel there, don't you? With adders between pairs X,Y and W,Z and between X,Z. That's how the hardware can do a DOT4 per clock.

Oh. I still was thinking CPU, I really have to change my thinking a bit now. I understand now that basically x,y,z,w,t are a little bit like an ALU-pool that can be connected/configured by the VLIW-instruction in quite a flexible way. The DOT4 instruction (for example) is not simply a serial product of it's inner products, it's really an instruction to create a specific configuration of the network of ALU nodes to accomplish the DOT4 not only in 1 clock in-order but also more exact. Right so far?
I think I still thought out-of-order, means I thought you get the same throughput of a single DOT4 instruction with multiple equivalent MUL/ADD instructions (using distinct outputs though).
Okay, my brain fires up, better wait for confirmation (GPUs == in-order?). Later the hailing of ATI's x,y,z,w,t concept. No wonder Fermi has so monstrous dimensions ...