AMD: R9xx Speculation

[racca]

TSMC's 40nm is the half node optical shrink as is 28 nm. They're taking the same approach for the next node and skipping straight to 20nm.
http://eetimes.com/news/latest/showArticle.jhtml?articleID=224400067

You couldn't have been more wrong. "Nodes" are just a way to focus foundries' investments.
A full-node implement reqiures new development tools, new litho-tools, new materials, etc. That's exactly what TSMC did with 40nm and 28nm. So it IS full node by all definition at least for 28nm. -- 32nm@TSMC was an optical shrink of 40nm it had no relationship with 28nm while it was planned.

Let's put it this way: would you call Phenom II a new uArch, just because it's released around the same time as Core i7 which has a considerably new uArch?

On the other hand most other foundries develop new tools at 32nm node then do an optical "shrink" with almost the same technology.

 

[psolord]

http://www.semiaccurate.com/2010/04/21/atis-southern-islands-tapes-out/

A hybrid gpu, having some parts of the current gen and some stuff from the next gen? Can someone speculate on that? What could change and what could remain the same?

Also is this thread about the Northern Islands or the Southern Islands? Which one is R9XX?

Personally, I am under the impression that both Nvidia and ATI can only do so much with any given amount of transistors, so if they want to really increase performance, they need to increase transistor count accordingly. So if it is 40nm I wouldn't expect anything great from it unless they decided to make it bigger.

As for the memory bandwidth that was discussed earlier, I think that ATI would be OK if they just used faster GDDR5 modules for now. If I am not mistaken, there are already 7Gbps modules available or about to be. So these modules would give 40% more bandwidth for S.I. and I don't really believe that S.I would give anywhere close to 40% more performance, in comparison to Cypress! :S

 

[neliz]

performance for the 6700 model is targeted 20% over the GTX480. tape-out was a bit longer ago than just recently. FLOP throughput on the 6800 might be 4 times faster

 

[CRoland]

performance for the 6700 model is targeted 20% over the GTX480. tape-out was a bit longer ago than just recently. FLOP throughput on the 6800 might be 4 times faster

Do you mean they more than double everything? Wouldn't that result in a huge die on 40nm?

 

[psolord]

performance for the 6700 model is targeted 20% over the GTX480. tape-out was a bit longer ago than just recently. FLOP throughput on the 6800 might be 4 times faster

What now?

Are you saying that the hypothetical 6770 should be 20% faster than the GTX 480?

And what about that "4 times faster"? The 6800 would be 4 times faster in FLOP throughput than the 6700 or the GTX 480? in SP or DP?

If the 6800 is 4 times faster in FLOP count than the 6700, what would that mean in terms of rendering performance? What I am thinking here is that the 5770 had only everything half of what the 5870 had,no?

=====================

Now that I am thinking of it, maybe I was a bit hasty of what I said about transistor counts. rv 870 is about twice as big than rv770 and still it did not give 100% more performance. It was closer to 60%. So there may be something wrong here that can be fixed.

On the other hand, nvidia's GF100 was indeed 50% bigger (always talking about transistor count) and still it gave a full 50% more performance or more.

So yes, actually there may be quite some room for improvement. But I cannot see that 5770 being 20% faster than GTX 480!

 

[rpg.314]

A hybrid gpu, having some parts of the current gen and some stuff from the next gen? Can someone speculate on that? What could change and what could remain the same?

My guess would be setup/raster/tess/caches/PCIe3(?)/12 monitor supportgrin. IOW, the uncore.

Also is this thread about the Northern Islands or the Southern Islands? Which one is R9XX?

I think whatever comes out of AMD next is R9xx.

Personally, I am under the impression that both Nvidia and ATI can only do so much with any given amount of transistors, so if they want to really increase performance, they need to increase transistor count accordingly. So if it is 40nm I wouldn't expect anything great from it unless they decided to make it bigger.

My guess would be higher clocks instead of higher core counts.

As for the memory bandwidth that was discussed earlier, I think that ATI would be OK if they just used faster GDDR5 modules for now. If I am not mistaken, there are already 7Gbps modules available or about to be. So these modules would give 40% more bandwidth for S.I. and I don't really believe that S.I would give anywhere close to 40% more performance, in comparison to Cypress! :S

Not sure how cheap it would be. IMHO, that jump should be easier to make than the one to GDDR5+.

 

[rpg.314]

On the other hand, nvidia's GF100 was indeed 50% bigger (always talking about transistor count)

No.

 

[CRoland]

Now that I am thinking of it, maybe I was a bit hasty of what I said about transistor counts. rv 870 is about twice as big than rv770 and still it did not give 100% more performance. It was closer to 60%. So there may be something wrong here that can be fixed.

It's around 80% on average with current drivers where it matters. (I.e. GPU limited <60 FPS.) But, yeah, could be better. Some things like memory bandwidth were not doubled, though.

On the other hand, nvidia's GF100 was indeed 50% bigger (always talking about transistor count) and still it gave a full 50% more performance or more.

50% more than 5870? Hardly. Usually less than 20%, though I'm not sure about the average.

 

[GZ007]

Personally, I am under the impression that both Nvidia and ATI can only do so much with any given amount of transistors, so if they want to really increase performance, they need to increase transistor count accordingly. So if it is 40nm I wouldn't expect anything great from it unless they decided to make it bigger.

They could go the other way. Make something like two 2bilion transistor dies on the same 40nm. If they could combine multiple dies on a single substrate than the 40nm shouldnt be a problem. (and maybe reach the 4x flop troughput of gtx480)
Or a xbox360 like xenos gpu-s daughter die with the edram and ROP-s could be caled uncore change (and also divide some transistor into two dies)

 

[CRoland]

They could go the other way. Make something like two 2bilion transistor dies on the same 40nm. If they could combine multiple dies on a single substrate than the 40nm shouldnt be a problem. (and maybe reach the 4x flop troughput of gtx480)

How about power consumption? Two dies would require double the power, all things equal.

 

[psolord]

Uh yes, I see my mistake.I thought that the GT200 was a 2 BIO chip as well. I confused it with the Cypress.

So all in all, both GF100 and Cypress are double than what their ancestors where and they both gave about 60%. Actually ATI's increased transistor count gave a tad more performance than what Nvidia's did.

So I take back that there is room for improvement, based just on that. There could still be room for improvement, but I cannot see that 6700 series being 20% faster than the GTX 480. At least not at 40nm!

 

[Mindfury]

在从5D减少到4D后SP总数非常接近1600个的情况下,9xx的性能我估计也就是比8xx高15~20%左右,核心DIE也大15~20%左右.
nApoleon 发表于 2010-4-22 14:43

Reduction from 5D to 4D after the SP is very close to the total number of 1,600, 9xx performance I guess that is higher than the 8xx 15 ~ 20%, the core DIE has 15 to 20% of large

这个我之前已经说过了,R9xx在曲面细分上会有加强,幅度不小
nApoleon 发表于 2010-4-22 09:11

I have said this before, R9xx essellation will be strengthened, not a small margin

 

[rpg.314]

Reduction from 5D to 4D after the SP is very close to the total number of 1,600, 9xx performance I guess that is higher than the 8xx 15 ~ 20%, the core DIE has 15 to 20% of large

So about 400mm2 die. But 4D clusters suggest 25SIMD's. I wonder how they have managed to make to two parallel modules referred here then

http://forum.beyond3d.com/showpost.php?p=1414636&postcount=424

 

[Mindfury]

But 4D clusters suggest 25SIMD's. I wonder how they have managed to make to two parallel modules referred here then

http://forum.beyond3d.com/showpost.php?p=1414636&postcount=424

nApoleon said R9xx chip would have less stream processors than Cypress.

So it could be 24 SIMD or 20 SIMD.

 

[leoneazzurro]

Maybe there will be a separate clock for shaders? Something like the "hot clock" and "hot clock/2" as in Fermi?

 

[mczak]

Reduction from 5D to 4D after the SP is very close to the total number of 1,600, 9xx performance I guess that is higher than the 8xx 15 ~ 20%, the core DIE has 15 to 20% of large

A reduction from 5D to 4D doesn't fit well with the "basically uncore changes" of SI. Or was that for NI? That gets confusing...

So all in all, both GF100 and Cypress are double than what their ancestors where and they both gave about 60%. Actually ATI's increased transistor count gave a tad more performance than what Nvidia's did.

Not really if you factor in that not all of Fermi's transistors are enabled. Given the increased complexity (with the distributed setup and all) and flexibility, perf/transistor sounds reasonable compared to GT200 (though of course GT200 was bad in comparison to G92 in that area). rv870 OTOH given the much smaller changes is disappointing by that measure, but it still has an advantage there so it doesn't really matter much.

So I take back that there is room for improvement, based just on that. There could still be room for improvement, but I cannot see that 6700 series being 20% faster than the GTX 480. At least not at 40nm!

Well, 20% would be like a 30% improvement from HD5870. I could very well see that happen, don't forget the story where AMD planned rv870 to be quite a bit bigger initially until they decided to scratch some parts. With all things learned until now (and 40nm still improving at TSMC) somewhat bigger chip (like ~20% bigger) might be easily feasible.

 

[Mindfury]

A reduction from 5D to 4D doesn't fit well with the "basically uncore changes" of SI. Or was that for NI? That gets confusing...

Maybe SI=NI on 40nm node...:smile:

 

[psolord]

Not really if you factor in that not all of Fermi's transistors are enabled. Given the increased complexity (with the distributed setup and all) and flexibility, perf/transistor sounds reasonable compared to GT200 (though of course GT200 was bad in comparison to G92 in that area). rv870 OTOH given the much smaller changes is disappointing by that measure, but it still has an advantage there so it doesn't really matter much.

Well, 20% would be like a 30% improvement from HD5870. I could very well see that happen, don't forget the story where AMD planned rv870 to be quite a bit bigger initially until they decided to scratch some parts. With all things learned until now (and 40nm still improving at TSMC) somewhat bigger chip (like ~20% bigger) might be easily feasible.

Ah yes, the GTX 480 is not a full GF100. Well, once a full GF100 comes, we'll have a better picture, I guess.

My objection to what Neliz said is that he talked about the 6700 being 20% faster than the GTX 480. Not the 6800 series. Obviously I objected due to the naming scheme of which we know nothing about. I just based my objection on the current naming scheme, ie the 6770 should be the higher end of the mainstream cards which in turn should be something much less than the supposed 6870.

Now for the 6870 to be 20% faster than the GTX 480, that is perfectly understandable.

For all we know, ATI could release just a 6770/6750 on this bastard architecture, which could be based on a 2.4BIO transistor chip, and release the rest of the cards, ie the 68XX at 28nm with the new architecture and about 4 BIO transistor chip. That way, the naming and the 20% increase over the GTX 480, would fit nicely. It would also put a ZOMG scare factor to Nvidia, knowing their high end baby got beaten by the mainstream card of ATI's next series.

I don't know how "mainstream" a card would be, being faster than the 5870 and for how much would it sell. ATI's cards appear to be getting all the more expensive with every coming generation. The 3870 was 200 euros (launch prices here), the 4870 was 250 euros, the 5870 350 euros. WTF ATI?

 

[rpg.314]

Ah yes, the GTX 480 is not a full GF100. Well, once a full GF100 comes, we'll have a better picture, I guess.

My objection to what Neliz said is that he talked about the 6700 being 20% faster than the GTX 480. Not the 6800 series. Obviously I objected due to the naming scheme of which we know nothing about. I just based my objection on the current naming scheme, ie the 6770 should be the higher end of the mainstream cards which in turn should be something much less than the supposed 6870.

Now for the 6870 to be 20% faster than the GTX 480, that is perfectly understandable.

For all we know, ATI could release just a 6770/6750 on this bastard architecture, which could be based on a 2.4BIO transistor chip, and release the rest of the cards, ie the 68XX at 28nm with the new architecture and about 4 BIO transistor chip. That way, the naming and the 20% increase over the GTX 480, would fit nicely. It would also put a ZOMG scare factor to Nvidia, knowing their high end baby got beaten by the mainstream card of ATI's next gen.

I don't know how "mainstream" a card would be, being faster than the 5870 and for how much would it sell. ATI's cards appear to be getting all the more expensive with every coming generation. The 3870 was 200 euros (launch prices here), the 4870 was 250 euros, the 5870 350 euros. WTF ATI?

Neliz could have been referring to just the flops.

 

[psolord]

Neliz could have been referring to just the flops.

Yes he could, lol.

Still, for the 6800 to be 4 times faster in flops than the GTX 480, would mean that it should be twice as fast as the 5870, since the 5870 is already twice as fast in flops compared to the GTX 480. Again I find this to be difficult for 40nm within the Cypress die size limits. Now if ATI is willing to make a much bigger chip, it is a different story altogether. I only hope the 6870 would replace the 5870 in both stature and price!