What went wrong with Fermi: JHH

[Keysplayr]

Define "full DX11 title", please.

















HINT: there's no such thing.

Oh, I get it. A play off the "essentially" quote you never answered.

::sigh:: It's always a game for you.

Word games. How about, list all DX11 titles no matter how little the DX11 content.

Can you keep it real long enough to have a conversation without the games?
I can.

 

[n0x1ous]

Does anyone really know the prices Nvidia or AMD pay per wafer. Isn't Nvidia supposed to be TSMC customer #1 with the best prices? If that's the case and they are getting better pricing then its hard to say for sure that the extra die size compared to AMD makes it a business failure.

Also what a lot of people seem to miss about the die size relative to Cypress, is that GF100 was NOT designed as a gaming chip. There is a considerable amount of die space dedicated to GPGPU that doesn't benefit gaming performance, and considering JHH has now said Nvidia is primarily focused on parallel computing, that makes sense to a degree. When Nvidia looked at GF100 and thought about how to optimize for gaming, we got GF104 which is much more competitive to Cypress as far as die size to performance.

I think it will be interesting to see what Kepler's architecture looks like as Nvidia has to walk the line between gaming and GPGPU. Maybe someone with more knowledge on the business end can shed some light on this......I don't see Nvidia developing two separate architectures for gaming and GPGPU, but (depending on how valuable the gaming market is to them at this point) wouldn't it make sense for them to build a very scalable architecture that is modular in the sense that they can do what they did with GF100 to GF104, but instead of releasing the GF100 type chip for both markets, they use the original GF100 like chip for Tesla etc and simultaneously pull out the ECC DFP stuff that doesn't benefit games and release a top end beast like normally do. As in make a 550mm GF104 (what the GTX 480 should have been) That way they are maximizing die space utilization in both markets instead of each market having things that are wasted?

If every generation they are using one architecture for both markets and the company leans toward parallel computing, then there will be increasing die space dedicated to GPGPU and it would be difficult to compete with AMD on performance per watt/die size. That to me seems like the biggest reason that Fermi is only 10-15% quicker than Cypress despite the much larger die. If Nvidia can find a way to make two different chips off the same arch. that are more focused on their respective markets we could then see the return of what we were used to seeing out of Nvida > huge dies, but huge performance to match that size.

Is that financially feasible or is the discrete market revenue potential not worth the investment anymore at this point to them?

 

[RussianSensation]

On a business level, Fermi is a failure. Why? It requires >50% die size to be competitive in performance.

Yup, from an efficiency/gaming performance per watt and manufacturing cost perspective, it plain sucks :thumbsup: But because it has GPGPU functions as well, it was probably impossible to design a GPU as fast as Cypress for games and better at GPGPU tasks while being that small.

 

[ModestGamer]

Does anyone really know the prices Nvidia or AMD pay per wafer. Isn't Nvidia supposed to be TSMC customer #1 with the best prices? If that's the case and they are getting better pricing then its hard to say for sure that the extra die size compared to AMD makes it a business failure.

Also what a lot of people seem to miss about the die size relative to Cypress, is that GF100 was NOT designed as a gaming chip. There is a considerable amount of die space dedicated to GPGPU that doesn't benefit gaming performance, and considering JHH has now said Nvidia is primarily focused on parallel computing, that makes sense to a degree. When Nvidia looked at GF100 and thought about how to optimize for gaming, we got GF104 which is much more competitive to Cypress as far as die size to performance.

I think it will be interesting to see what Kepler's architecture looks like as Nvidia has to walk the line between gaming and GPGPU. Maybe someone with more knowledge on the business end can shed some light on this......I don't see Nvidia developing two separate architectures for gaming and GPGPU, but (depending on how valuable the gaming market is to them at this point) wouldn't it make sense for them to build a very scalable architecture that is modular in the sense that they can do what they did with GF100 to GF104, but instead of releasing the GF100 type chip for both markets, they use the original GF100 like chip for Tesla etc and simultaneously pull out the ECC DFP stuff that doesn't benefit games and release a top end beast like normally do. As in make a 550mm GF104 (what the GTX 480 should have been) That way they are maximizing die space utilization in both markets instead of each market having things that are wasted?

If every generation they are using one architecture for both markets and the company leans toward parallel computing, then there will be increasing die space dedicated to GPGPU and it would be difficult to compete with AMD on performance per watt/die size. That to me seems like the biggest reason that Fermi is only 10-15% quicker than Cypress despite the much larger die. If Nvidia can find a way to make two different chips off the same arch. that are more focused on their respective markets we could then see the return of what we were used to seeing out of Nvida > huge dies, but huge performance to match that size.

Is that financially feasible or is the discrete market revenue potential not worth the investment anymore at this point to them?

Translation.

there alot of really poorly written code for gpgpu use and the OS kernel are so underoptimized for this type of computing that the only way to even approach the capability is to add needless logic onto the gpu die.

BTW it hardly adress's the issue.

 

[n0x1ous]

Translation.

there alot of really poorly written code for gpgpu use and the OS kernel are so underoptimized for this type of computing that the only way to even approach the capability is to add needless logic onto the gpu die.

BTW it hardly adress's the issue.

And what exactly is the issue? Like Russian said above It would be tough for them to beat Cypress with same die size when they need to include the ECC DFP etc. Im not saying saying if they would have left all that out thats its a foregone conclusion that they would have beat Cypress on perf/die size, but certainly it would have been much closer as GF104 shows

 

[JimmiG]

I think it was a mistake to use the same GPU in both their gaming products and their GPGPU products. As has been said (and certainly Nvidia must have realized this, too), there are a lot of transistors on the GF100 that take up space and draw power without adding anything useful when running games.

GF104 has only 2/3's the transistor count, a much smaller die area and it draws 65W less than the GTX470 (100W less than GTX480), yet it could easily be as fast as the GTX470. The GTX460 isn't, but a GTX460 with all 384 cores enabled and 800 MHz clock speeds would be...

Maybe they just didn't have time to pull out all the non-gaming GPGPU fluff from the GF100 before launching their gaming GPU's, as it was already late to the market..

 

[RussianSensation]

JimmiG, I think it's very expensive to design a standalone GPGPU product at this time given that this industry and related software support are still in their infancy. If general purpose computing picks up in the next 5-10 years, I could see them going that route. At this time, it already costs $1-2B to develop a new GPU architecture. I doubt NV has the funds to do 2 separate revisions for 2 different markets at this time.

 

[n0x1ous]

Russian, so do you think that what we have seen as far as performance differences between Nvidia and AMD this generation will be indicitive of the future as long as Nvidia has to make an architecture that compromises gaming for GPGPU and vice versa?

Or can they just make a 500mm^2 GF104 type chip without the GPGPU stuff as the launch flagship next time instead of the cutdown performance version that is GF104 today?

 

[RussianSensation]

Not sure about the performance differences going forward. NV and ATI have traded performance many times in the last 10 years. However, unless NV changes their strategy, they will continue to make larger sized chips vs. small-die strategy of AMD. That is almost certain unless one of the firms changes their direction.

GF100/GF104/106 are decent chips from a performance perspective, but are not very efficient as a result of the bloated architecture. NV has noted that through new architectures and more optimized software they will attempt to improve performance per watt. ATI will continue to lead in this area in the foreseeable future imo simply because their strategy is to design a smaller die and ship 2x smaller die chips for high end cards vs. NV's strategy which is to scale a very large die down towards lower price segments. 2 Completely different strategies. So far AMD has been more successful.

 

[n0x1ous]

I hope for Nvidia's sake that the GPGPU stuff works out well for them. I do use a CUDA transcoding app myself and would love to be able to use it for video editing as well, but can't justify spending more for Premier Pro when elements works fine for what I need.

Maybe eventually they will have enough cash to devote 2-4 billion and make more of a distinction between the markets. I'm curious what AMD spends on R&D for an architecture in comparison. Certainly its less then Nvidia as AMD only focuses on gaming performance for the most part...

Isn't Nvidia bigger than even the combined AMD+ATI....I thought their market cap was like 7 billion vs like 5, but I could be wrong....

 

[tweakboy]

What went wrong is the 335watt demand for their single GPU 480 card.

ATI is 200watts or less on their dual GPU 5970 .

That is what went wrong.

 

[Keysplayr]

What went wrong is the 335watt demand for their single GPU 480 card.

ATI is 200watts or less on their dual GPU 5970 .

That is what went wrong.

Wrong,

and

Wrong again.

Why dude?

 

[Madcatatlas]

Right and right again

why?

 

[Keysplayr]

Right and right again

why?

I feel like I'm in nursery school. If you can explain to me how Tweakboy was right, I'll give you a cookie.

 

[notty22]

The 5970 uses two gpu's that are rated for 181 watts when used in the 1gb 5870.
A 2gb 5870 is rated at 220.
The 5970 is able to 'achieve' its 300* watt tdp because the clocks are lowered on low leakage 'select' dies.
Imho, ATI will be hard pressed to make a better flagship dual gpu card.
I say this because, the gpu's definitely scale and make power to 925mhz. A stock 5970 is tested usually at its stock speeds of 725. Giving enthusiasts lots of room to overclock, if they can also deal with the heat power issues of the beast.

*http://www.amd.com/us/products/desk...5970/Pages/ati-radeon-hd-5970-overview.aspx#2

 

[Lonyo]

I feel like I'm in nursery school. If you can explain to me how Tweakboy was right, I'll give you a cookie.

Well, he's tweakboy, so you have to give him some leeway.
The HD5870 does use less power than the GTX480 though (20w according to Xbitlabs, 40w according to Anandtech).
While he has exaggerated with his figures, the overall point is valid, the GTX480 is power hungry, even moreso than the dual GPU HD5970.

http://www.xbitlabs.com/articles/video/display/gigabyte-gf-gtx400_6.html#sect0
http://www.anandtech.com/show/2977/...x-470-6-months-late-was-it-worth-the-wait-/19

 

[Keysplayr]

Well, he's tweakboy, so you have to give him some leeway.
The HD5870 does use less power than the GTX480 though (20w according to Xbitlabs, 40w according to Anandtech).
While he has exaggerated with his figures, the overall point is valid, the GTX480 is power hungry, even moreso than the dual GPU HD5970.

http://www.xbitlabs.com/articles/video/display/gigabyte-gf-gtx400_6.html#sect0
http://www.anandtech.com/show/2977/...x-470-6-months-late-was-it-worth-the-wait-/19

Did you think there was a possibility that I didn't know this info?

I was addressing Madcatatlas, wondering why he mocked my post.

Tweakboy was obviously incorrect on both counts. Madcat apparently believed him.