AMD: R9xx Speculation

[Novum]

Reorganization of shader ALUs doesn't result in a "new shader model". I suspect the entire post to be fake.

The 98.5% efficiency claim is pretty unlikely as well. If you can on average fill 3.5 of 5 VLIW slots on the current architecture, that doesn't mean you can fill 3.5 of 4 on the new one. You can easily construct examples where this isn't the case.

 

[rpg.314]

http://www.chiphell.com/thread-121500-1-1.html

Caicos (HD6350):
Vantage P:17XX

More:
http://i52.tinypic.com/2r5w7xc.jpg

http://i55.tinypic.com/1zbzaex.jpg

http://i55.tinypic.com/34hji2h.jpg

Is this one going to be a bottom up launch?

 

[neliz]

Is this one going to be a bottom up launch?

Don't hold your breath for that.

 

[Zaphod]

I can't find 98.5% mentioned anywhere on that post. Also, not sure how he came with that number.

It's probably been devined from this paragraph:

Now just like this freakin' toll booth analogy i just created, the Evergreen shader was the old 5 lane booth, and the N. Islands shader is the newly constructed 4 lane booth. It's more efficient, and while the 5-lane booth can pass 10,000 cars per hour, the 4-lane booth accepts slightly fewer- 9,850 cars per hour while using ~80% of the physical footprint. It's a more efficient shader design.

 

[mczak]

The 98.5% efficiency claim is pretty unlikely as well. If you can on average fill 3.5 of 5 VLIW slots on the current architecture, that doesn't mean you can fill 3.5 of 4 on the new one. You can easily construct examples where this isn't the case.

It would be possible efficiency is increased over 100% for some scenarios depending on how the 4 remaining units incorporated the capabilities of the T unit (assuming this is what is gone).
Still, the numbers sound a bit too impressive for me too, but I wouldn't rule it out completely.

 

[Jawed]

Week 16

 

[Man from Atlantis]

Week 16

its 1028, so i guess 28th week..

 

[mczak]

Caicos (HD6350):
Vantage P:17XX

That would be an improvement of about 30% over HD5450. Not mindblowing (still can't touch GT220 for example) but if the chip is only ~10% bigger why not?
Might mean that indeed shader units doubled (or rather, if the 4D rumors are true, 80% more in the end only) and the simds are now 16 wide like on all other cards.

 

[Jawed]

What is AMD's Northern Islands?

http://www.semiaccurate.com/2010/09/06/what-amds-northern-islands/

Lets start out by saying we got something very wrong at SemiAccurate, the Southern Islands name. No, SI is real, it is the 28nm version of the chip that we have been calling Northern Islands for a few months now. The dates, functions and all the rest we had correct, we just reversed the family names.

I once pondered whether the names are reversed

Back to the NI family, what are they? Well, that part is easy enough, they are a serious re-do of the Evergreen family. The biggest change is in the shaders, they have gone from a 4 simple + 1 complex arrangement to a 4 medium complexity arrangement.

 

[Silent_Buddha]

Anyone care to comment on this? Particularly the SP redesign. The author claims a reduction to 4-wide SPs which allows ATi to fit an additional 25% SPs into each chip. If true, I find it a bit odd as it would give them slightly less mathematical power at the same die size. Cypress: 320x5 = 1600, Cayman: 320(*1.25)x4 = 1600. Seems a bit odd to me, unless there's more to the story or perhaps more SIMDs will be used, increasing the die size.

Well, assuming that rumor is true. ~98.5% of the performance in ~80% of the space still leads to an overall performance/mm^2 increase of ~24.5%. Which to me is a rather large increase in performance at the same die size. Where before you had say 800 shaders each running at 1x performance (averaged) you'd now have ~1000 shaders each running at 0.985x performance (averaged) in roughly the same die area.

Regards,
SB

 

[MarkoIt]

Well, if all what Charlie said is true (25% up for the perf/mm^2), in 380 mm^2 they could fit 50% more performance than Cypress... not far from the 35% percent leaked out for both Cayman and Caicos over Evergreen series

 

[Erinyes]

We've seen a Caicos board(second gen DX11 part) before we've seen a GF108 board. That tells you everything. If they've got a working board already, release cant be that far off.

 

[DavidGraham]

Then I suppose the 4D shaders is true after all , it is about time I think .. we should expect a performance increase of about 35~50% (@1920SP) , provided that the rest of the chip (textures and front end ) undergo an exponential increase in count as well .

 

[Sontin]

We've seen a Caicos board(second gen DX11 part) before we've seen a GF108 board. That tells you everything. If they've got a working board already, release cant be that far off.

You missed this: http://forum.beyond3d.com/showpost.php?p=1458653&postcount=6434

 

[Mianca]

Well, assuming that rumor is true. ~98.5% of the performance in ~80% of the space still leads to an overall performance/mm^2 increase of ~24.5%. Which to me is a rather large increase in performance at the same die size. Where before you had say 800 shaders each running at 1x performance (averaged) you'd now have ~1000 shaders each running at 0.985x performance (averaged) in roughly the same die area.

Regards,
SB

Yeah, or even more vivid: A 1280SP (320*4) Barts chip could very well offer similiar clock-for-clock shader performance to a 1600SP (320*5) Cypress chip while keeping die area around 280mm2

They'll have to artifically restrain clock speeds on Barts quite a bit in order to make the HD 6770 card perform along the lines of the old HD 5830 card (which is where they SHOULD place it from a marketing perspective): Given that HD 5830 is on average about 20-25% slower than HD 5870, I'd expect the HD 6770 to keep the core clock at about 650-700Mhz.

An HD 6970 Antilles card with 2xBarts @ "full" 850-875 Mhz (I guess with Charlie's info concerning Cayman's die size probably ending up a few mm2 below 400mm2 we can lay Cayman-based Antilles rumours to rest) would trump the current HD 5970 card by about 20%

 

[rpg.314]

Yeah, or even more vivid: A 1280SP (320*4) Barts chip could very well offer similiar clock-for-clock shader performance to a 1600SP (320*5) Cypress chip while keeping die area around 280mm2

They'll have to artifically restrain clock speeds on Barts quite a bit in order to make the HD 6770 card perform along the lines of the old HD 5830 card (which is where they SHOULD place it from a marketing perspective): Given that HD 5830 is on average about 20-25% slower than HD 5870, I'd expect the HD 6770 to keep the core clock at about 650-700Mhz.

An HD 6970 Antilles card with 2xBarts @ "full" 850-875 Mhz (I guess with Charlie's info concerning Cayman's die size probably ending up a few mm2 below 400mm2 we can lay Cayman-based Antilles rumours to rest) would trump the current HD 5970 card by about 20%

Why does the idea of Antilles = 2xCayman sound so unrealistic to you?

 

[UniversalTruth]

You are continuing with your guesstimations...
Antilles is two times Cayman. May you accept it?

 

[UniversalTruth]

...An HD 6970 Antilles card with 2xBarts @ "full" 850-875 Mhz (I guess with Charlie's info concerning Cayman's die size probably ending up a few mm2 below 400mm2 we can lay Cayman-based Antilles rumours to rest) would trump the current HD 5970 card by about 20%

Sorry, but you would you be so kind to explain to all of us how this would happen. If Barts is slower or equal in performance to Cypress?

 

[mczak]

Yeah, or even more vivid: A 1280SP (320*4) Barts chip could very well offer similiar clock-for-clock shader performance to a 1600SP (320*5) Cypress chip while keeping die area around 280mm2

They'll have to artifically restrain clock speeds on Barts quite a bit in order to make the HD 6770 card perform along the lines of the old HD 5830 card (which is where they SHOULD place it from a marketing perspective): Given that HD 5830 is on average about 20-25% slower than HD 5870, I'd expect the HD 6770 to keep the core clock at about 650-700Mhz.

If Barts really is 1280SP AND they perform like Cypress 1600SP, given there should be other improvements Barts would be faster than Cypress at the same clock. So, if that is smaller than Cypress I can see no reason to keep Cypress around - it should be cheaper to produce after all. If they'd want to have some version which is as fast as HD5830 then I'd guess AMD could just cripple it the same way with the somewhat broken half-disabled rops (though I guess maybe this time they got it right and it will have less of a performance impact, also certainly there is nothing wrong with disabling some simds). That would be a killer chip though against GF104...

 

[Jawed]

If this VLIW-4 rumour is true then Barts with 1280 lanes would have the same TMU count as Cypress. Also, games don't seem to be ALU bound, so 1280 versus 1600 is probably not going to be much of an issue.

HD5870 game performance for $200 would be pretty cool.