AMD: R9xx Speculation

[rpg.314]

Barts is 1/3 of GF100, i.e. about 175mm²

That's almost the same size as juniper. I wonder what was "fixed"?

 

[UniversalTruth]

If the die size is really in that ballpark and the names of the products truely related to die size, then 6800 is not fair and only ugly marketing trick in order to confuse people and try to sell more of these mainstream sh...s...

I hope it can help to stop the pointless discussion...

With only 2 exceptions, names of the ATi's products launched in last 7 years, always corresponded to die size:

101 - 150mm² => x6xx

• RV830 (Radeon HD5500): 104 mm²
• RV635 (Radeon HD3650): 120 mm²
• RV530 (Radeon X1600): 150 mm²
• RV630 (Radeon HD2600): 150 mm²
• RV730 (Radeon HD4600): 150 mm²

151 - 190mm² => x7xx

• RV410 (Radeon X700): 156 mm²
• RV840 (Radeon HD5700): 170 mm²

191 - 340mm² => x8xx

• RV670 (Radeon HD38x0): 192 mm²
• R430 (Radeon X800XL): 240 mm²
• RV770 (Radeon HD4800): 256 mm²
• R420 (Radeon X800XT): 281 mm²
• RV790 (Radeon HD4890): 282 mm²
• R520 (Radeon X1800): 288 mm²
• RV870 (Radeon HD5800): 331 mm²

341 - 500mm² => x9xx

• R580 (Radeon X1900XT): 342 mm²
• R580+ (Radeon X1950XT): 34x mm²
• R600 (Radeon HD2900): 420 mm²

The only two exceptions in this system were two "pipe-cleaner" parts, which were used for testing of new manufacturing process and because of that, they were significantly smaller, than other GPUs of that family: RV570 (80nm), RV740 (40nm).

We expect Cayman to be sized around 400 mm² and Barts under 300 mm². What name should these two parts use? I think AMD isn't going to change anything. Maybe the users misinterpretted naming convetions.

 

[Jawed]

Compare:

http://thinklaptops.net/wp-content/uploads/2010/09/Sapphire-Radeon-HD-5770-Back-e1284389922765.jpg

and

http://i51.tinypic.com/t89dg1.jpg

Barts looks significantly larger than Juniper.

 

[Wirmish]

Note: Dies cropped when each PCB was on the same scale.

 

[jimbo75]

If these rumours are true AMD is making a mistake I fear. If the 6850 and 6870 are slower than the 5850 and 5870, they are also going to make nVidia's cards look better as well. I can't see why AMD would want their 6800's to be slower than say the 470 and not a lot faster than the 460. That's playing into nV's hands.

There really better be a good reason for these rumours being true.

 

[UniversalTruth]

@Wirmish: Please explain. Where do you get the idea that each pixel on those three images corresponds to exactly the same amount of space in real world?

 

[Man from Atlantis]

Compare:

http://thinklaptops.net/wp-content/uploads/2010/09/Sapphire-Radeon-HD-5770-Back-e1284389922765.jpg

and

http://i51.tinypic.com/t89dg1.jpg

Barts looks significantly larger than Juniper.

here a quick match

 

[rpg.314]

If these rumours are true AMD is making a mistake I fear. If the 6850 and 6870 are slower than the 5850 and 5870, they are also going to make nVidia's cards look better as well. I can't see why AMD would want their 6800's to be slower than say the 470 and not a lot faster than the 460. That's playing into nV's hands.

There really better be a good reason for these rumours being true.

The new 68xx needn't be a disaster if they cost half as much as a 58xx and are in the same ballpark.

 

[UniversalTruth]

The new 68xx needn't be a disaster if they cost half as much as a 58xx and are in the same ballpark.

Dreams are nice thing but with this size and this performance I don't see how exactly it will be cheaper that much...


By the way. I expected performance in Cypress' ballbark but with Juniper's die size.

 

[Pete]

Those numbers are weird. 6850 seems on par with 5850, yet 6870 falls way short of 5870 (i.e. reaches ~ GTX 470 performance).

Keep in mind that newer drivers boosted 5870 scores by almost 10%. If the 5850 benefits equally, that would bring it to ~7100, which splits the difference between the two purported Barts scores.

Edit: Just because I have no idea of what parts of a card affect Vantage perf, a bigger table using newer drivers:

Spoiler

[b]        X	core	SPs	GFLOPs	GB/s	[/b]
5850	6566	725	1440	2088	128.0	TechGage1
5870	8213	850	1600	2720	153.6	9.10, 3.33GHz
		25.1%	17.2%	11.1%	30.3%	20.0%	
						
5850	7005	725	1440	2088	128.0	[url=http://techgage.com/article/nvidia_geforce_gts_450_-_the_super-affordable_fermi/8]TechGage 450 review[/url]
5870	8816	850	1600	2720	153.6	10.8, 4GHz
	25.9%	17.2%	11.1%	30.3%	20.0%	
						
5850	7483	725	1440	2088	128.0	[url=http://www.xbitlabs.com/articles/video/display/nvidia-geforce-gtx-460_15.html#sect1]Xbit 460 review[/url]
5870	8870	850	1600	2720	153.6	10.6, 3.33GHz
	18.5%	17.2%	11.1%	30.3%	20.0%	
						
	[b]X	Gpix/s	Gtex/s	GFLOPs	GB/s	[/b]
5850	7483	23.2	52.2	2088	128.0	[url=http://www.xbitlabs.com/articles/video/display/nvidia-geforce-gtx-460_15.html#sect1]Xbit 460 review[/url]
5870	8870	27.2	68.0	2720	153.6	10.6, 3.33GHz
	18.5%	17.2%	30.3%	30.3%	20.0%	
						
5830	6008	12.8	44.8	1792	128.0	
5850	7483	23.2	52.2	2088	128.0	
	24.6%	81.3%	16.5%	16.5%	0.0%	
						
	[b]X	MHz	Gbps	ALUs[/b]		
6850	6549	775	4000	800		leaked #s
6870	7538	900	4200	960		i7-920 2.93
	15.1%	16.1%	5.0%	20.0%		
					[b]TMUs[/b]	
5850	7483	725	4000	1440	56	Xbit #s
6870	7538	900	4200	960	48	specs taken from [url=http://image208.poco.cn/mypoco/myphoto/20100915/13/557742492010091513152203.jpg]here[/url]
	0.7%	24.1%	5.0%	-33.3%	-14.3%

If Barts has such an ALU deficit, could a revamped front end account for how close it gets to 5850? Would Vantage even demonstrate that difference?

 

[trinibwoy]

Why in the world would you expect that on the same process node?

 

[UniversalTruth]

Because of the optimised architecture. I think it's time for AMD to do something in that direction, not just adding raw power...

 

[AnarchX]

here a quick match

http://img100.imageshack.us/img100/7946/junipervsbartsfixed.png

Did you believe this analysis offers more than the information that Barts has a 256-Bit MC, which brings the need for a bigger package?

Maybe AMD laid the transistor a bit looser and aims for >1GHz on XT-SKUs?

 

[rpg.314]

A ~190mm2 chip *might* be too small for a 256 bit bus. I remember rv770 being pad limited for a 256 bit bus @260mm2. Not to mention that the memory bus is likely to run much faster than rv770.

 

[Wirmish]

@Wirmish: Please explain. Where do you get the idea that each pixel on those three images corresponds to exactly the same amount of space in real world?

I take care to correct perspective, tapeze, tilt, etc, of the 3 PCB, until they are on the same scale, just like Man from Atlantis did.
I also measured the 3 PCB height to ensure that the 3 images are on the same vertical scale.
Of course, there is no link between the pixels on those 3 images and the real world, since each monitor has a different resolution and pitch.
But now that they are at the same scale, it's easier to compare them.

 

[AnarchX]

A ~190mm2 chip *might* be too small for a 256 bit bus. I remember rv770 being pad limited for a 256 bit bus @260mm2. Not to mention that the memory bus is likely to run much faster than rv770.

If we are looking at rumors about 1000-1050MHz memory clock on Barts they might have downgraded the MC speed to save die space?

 

[Wirmish]

A ~190mm2 chip *might* be too small for a 256 bit bus. I remember rv770 being pad limited for a 256 bit bus @260mm2.

Did you believe this analysis offers more than the information that Barts has a 256-Bit MC, which brings the need for a bigger package?

Bart can have a 256-bit bus... or a 128-Bit bus + GDDR5 @ 6 GHz.
Bart can have a 256-bit bus with a 192mm² (RV670) die size.
Bart can have a small die (~Juniper), but a big package (~Cypress).
Bart can be Cayman, and Cayman can be Bart.
Who knows ?

 

[UniversalTruth]

...Who knows ?

Charlie knows:

Barts is smaller than Cayman. Cayman is the biggest chip out there, and Antilles is a dual Cayman. Consider this a 'know', not a guess.

-Charlie

http://semiaccurate.com/forums/showthread.php?p=75038#post75038

 

[no-X]

A ~190mm2 chip *might* be too small for a 256 bit bus. I remember rv770 being pad limited for a 256 bit bus @260mm2. Not to mention that the memory bus is likely to run much faster than rv770.

RV770 had
- 3x display output (2x digital, 1 analog TV-out)
- 2x CrossFire port
- 1x sideport

Cypress have
- 6x display output
- 2x CrossFire ports
- no sideport

Barts has (probably)
- 5x display output
- 1x CrossFire port
- no sideport

We know, that they ditched side-port because of die size. But Barts (compared to Cypress) lacks 1 display output and 1 CrossFire port. It is possible, that Barts perimeter isn't big enough to fit all these ports. I think that die size around 210mm² is quite possible. Die size around 200mm² could mean, that Barts isn't 2/3 of Cayman, but only 50%(?)

 

[Alexko]

A ~190mm2 chip *might* be too small for a 256 bit bus. I remember rv770 being pad limited for a 256 bit bus @260mm2. Not to mention that the memory bus is likely to run much faster than rv770.

There were probably other factors at play with RV770. RV670 was fine with a 256-bit bus on a 192mm² die, and on a 55nm process at that.