6870 VS GTX460 Die sizes and price/performance

[bryanW1995]

I'll be able to tell you in ~ 38 hrs

 

[GaiaHunter]

Thread is long and well off-topic.

Can anyone tell me whether or not the numbers in this BSN article are considered "accurate"?


http://www.brightsideofnews.com/new...lands-vs-nvidia-fermi-die-sizes-compared.aspx

The Barts die size seems to be that according to die measurements - just check stuff like B3D forums.

Cayman I haven't seen any die pictures, but I guess that goes with the "Barts being 2/3 of Cayman".

 

[AtenRa]

If Cayman will be double the Barts (Shaders, Texture Units) except the memory controllers and the ROPs it could be around 400mm2.

Cayman ~ 400-420mm2
2240 Shaders with 96 Tex Units with 256-bit memory controllers and 32 ROPs

Barts = 255mm2
1120 Shaders with 48 Tex Units with 256-bit memory controllers and 32 ROPs

 

[HurleyBird]

If Cayman will be double the Barts (Shaders, Texture Units) except the memory controllers and the ROPs it could be around 400mm2.

Cayman ~ 400-420mm2
2240 Shaders with 96 Tex Units with 256-bit memory controllers and 32 ROPs

Barts = 255mm2
1120 Shaders with 48 Tex Units with 256-bit memory controllers and 32 ROPs

Or:

Cayman ~ 400-420mm2
2240 Shaders with 96 Tex Units with 256-bit memory controllers and 64 ROPs

or

Cayman ~ 420+mm2
2240 Shaders with 96 Tex Units with 512-bit memory controllers and 64 ROPs

or, the most interesting possibility IMHO:

Cayman ~ 400-420mm2
2240 Shaders with 96 Tex Units with 256-bit memory controllers w/ ECC and 32/64 ROPs

If Cayman has ECC, it could be going up against Fermi in HPC in a big way. Also explains why the memory controller remains 256-bit at the larger die size.

 

[busydude]

If Cayman has ECC, it could be going up against Fermi in HPC in a big way. Also explains why the memory controller remains 256-bit at the larger die size.

Doesn't Cypress have a memory controller with ECC?

 

[tviceman]

Thread is long and well off-topic.

Can anyone tell me whether or not the numbers in this BSN article are considered "accurate"?


http://www.brightsideofnews.com/new...lands-vs-nvidia-fermi-die-sizes-compared.aspx

Fudzilla, the other most accurate rumor mill, is saying it's 255. There is also plenty of reports saying it's ~240. Since most of the rumors/leaks are in the same ballpark, it's likely right in this neighborhood, and we'll all probably know for sure on Friday.

 

[AtenRa]

Barts is 255mm2

 

[Lonyo]

Doesn't Cypress have a memory controller with ECC?

No, it has a memory controller than detects errors and resends the data, but it's not ECC.

http://www.anandtech.com/show/2841/12

 

[Martimus]

Fudzilla, the other most accurate rumor mill, is saying it's 255. There is also plenty of reports saying it's ~240. Since most of the rumors/leaks are in the same ballpark, it's likely right in this neighborhood, and we'll all probably know for sure on Friday.

I have to admit that I never would have thought I would hear the words "Fudzilla" and "most accurate" in the same sentence. Even though you are obviously being facetious, it is still very strange seeing that sentence written.

 

[RaistlinZ]

I hope Crossfire scaling is much improved with the 68xx series. Regardless, I'll just sit and wait patiently for the 6970 like a good little boy.

 

[tviceman]

I have to admit that I never would have thought I would hear the words "Fudzilla" and "most accurate" in the same sentence. Even though you are obviously being facetious, it is still very strange seeing that sentence written.

I'm glad you got the joke!

 

[psoomah]

Or:

Cayman ~ 400-420mm2
2240 Shaders with 96 Tex Units with 256-bit memory controllers and 64 ROPs

Cayman doesn't need to annihilate the 480, just soundly beat it ... with a die half the size. If you know ahead of time what your competition will be you can design to maximize profits.

Why wouldn't Cayman be an extension of the Barts philosophy ---> decide what performance is needed to beat a known target and squeeze it into the smallest die size possible.

That 'known target' is the key. By the time AMD was finalizing their NI for 40nm design, AMD had a pretty clear view of what Nvidia would be able to bring to the table in 2010, making it possible to design to specific performance points and achieve optimized die sizes.

Considering the performance AMD squeezed out of the 255mm2 Barts die, my question would be how big does the Cayman die need to be to meet the performance objective --> say a 10-20% increase over the 480.

Could they achieve it with a die small enough to still be pin compatible with cypress ... ~ 340mm2?

Maybe. Or maybe they just squeezed as much performance into the maximum die size that could remain pin compatible and let it go at that. 340mm would do that and meet the 1/3 ~ 2/3 ~ 3/3 stategy goals. That thirds strategy leaves a space for a x2 Barts card (2x2/3 = 4/3) and a x2 Cayman card 2x3/3 = 6/3 which could reprise the 6970 TDP strategy.

 

[AtenRa]

You have to understand that 6870 is a chopped Cypress, they chopped 6 SIMDs off (480 Shaders) and they simple installed a new Front End with a new generation Tesselator in the Graphics Engine. They then raised the core frequency to 900MHz in order to have the same performance with 5850.

The only thing AMD won with BARTS is die size.

1440 vs 1120 = 28,57% difference in Shaders
725MHz vs 900MHz = 24,13% difference in core frequency.
5850 has 2088 Gflops and 6870 has 2016 Gflops. Almost the same.

5850 Max Board Power 151 Watt
6870 Max Board Power 151 Watt

5850 die = 334mm2
6870 die = 255mm2

69xx series will not get the same performance increase as barts took, if they will just double the SIMDs (28 = 2240 Shaders) and the texture units (96).

 

[toyota]

Barts is 255mm2

how do we know that is correct because they don't even have the correct bandwidth listed for the 5850?

 

[AtenRa]

Yes the memory bandwidth for the 5850 is wrong but is right for 6870. I guess a typo.

 

[nyker96]

AMD's claim of 30% improvement on same die size is little fishy. I'd be very happy if they can get 15-20% on same die size. But that is why I'm anxious to see some real reviews popping up for hd6xxx, even if the claims are fake still get my blood boiling.

 

[RussianSensation]

Why wouldn't Cayman be an extension of the Barts philosophy ---> decide what performance is needed to beat a known target and squeeze it into the smallest die size possible.

That 'known target' is the key.

HD6000 GPU design was likely close to final before GTX480 even saw the light of day in March 2010. GPU designs take years. They don't just sit down and add another 500 million transistors because that's all they need to 'beat a GTX480'. The last 6 months are used to finalize clock speeds , and decide how much of the chip to disable to achieve good yields and good power consumption.

Plus, HD6000 series is going to compete with GF110 or whatever refresh NV will have. Therefore the target is actually unknown. AMD's design strategy is to maximize performance/watt and use dual-GPUs to compete with a single high-end NV card (that was their new strategy since HD3800 series). NV dropped the ball though and GTX480 got walked over by 5970. AMD's philosphy is small-die vs. large die for NV. I don't think AMD ever anticipated that their "small die" chip is going to be faster than NV's larger die. That's a side-effect of their amazing engineering.

 

[AtenRa]

http://www.tomshardware.de/foren/241537-41-verkaufe-radeon-6870#t17807


6870

hmm wrong thread ehehe

 

[tviceman]

HD6000 GPU design was likely close to final before GTX480 even saw the light of day in March 2010. GPU designs take years.

NO THEY DON'T! YOU ARE TALKING NONSENSE! Cayman was designed overnight back in April! Nvidia will now be FOREVER 6 months late with every future release because of FERMI!

 

[busydude]

No, it has a memory controller than detects errors and resends the data, but it's not ECC. http://www.anandtech.com/show/2841/12

Thanks for clearing that up. Appreciate it.

 

[tviceman]

http://www.tomshardware.de/foren/241537-41-verkaufe-radeon-6870#t17807


6870

I'm not sure what graphical settings (beyond what is shown) this leak is at, but here is a huge graph of all current cards with Unigine running at a slightly higher resolution on "high" settings. http://ixbtlabs.com/articles3/i3dspeed/0710/itogi-video-h2-wxp-1920-pcie.html I searched for a few minutes and could not find equal comparisons with several cards in the mix.

At a slightly higher resolution, the hd5870 is still faster than this leak shows the hd6870 to be at the lower resolution indicated. If this holds true among other DX11 benchmarks, Silverforce's statements about the hd6870 being faster in DX11 than hd5870 will turn out to be false. Improved performance? Yes. But not faster than Cypress.

 

[tijag]

I'm not sure what graphical settings (beyond what is shown) this leak is at, but here is a huge graph of all current cards with Unigine running at a slightly higher resolution on "high" settings. http://ixbtlabs.com/articles3/i3dspeed/0710/itogi-video-h2-wxp-1920-pcie.html I searched for a few minutes and could not find equal comparisons with several cards in the mix.

At a slightly higher resolution, the hd5870 is still faster than this leak shows the hd6870 to be at the lower resolution indicated. If this holds true among other DX11 benchmarks, Silverforce's statements about the hd6870 being faster in DX11 than hd5870 will turn out to be false. Improved performance? Yes. But not faster than Cypress.

If xbit labs uses an i7-920 @ 4.0ghz then no doubt that wouldn't be a fair comparison to this benchmark using a PhII.

Also, hopefully games aren't using the absurd overkill method of tessellation which doesn't really improve graphics, but causes a huge drain on performance.

 

[railven]

If xbit labs uses an i7-920 @ 4.0ghz then no doubt that wouldn't be a fair comparison to this benchmark using a PhII.

Also, hopefully games aren't using the absurd overkill method of tessellation which doesn't really improve graphics, but causes a huge drain on performance.

According to TV's link the 5870 got 34.5 versus the leaks 34.3

Thinks to factor in:
TV's link has no AA, the leak has 4xAA
TV's link has no ANIS, the leak has trilinear

I can't find the processor but chances are it isn't a Phenom II, X4, probably a faster Core i7 with an OC.

 

[railven]

Found this:

http://ixbtlabs.com/articles3/i3dspeed/0910/i0910-video-p1.html

Testbed configuration

• Intel Core i7-975 3340 MHz CPU
• ASUS P6T Deluxe motherboard on the Intel X58 chipset
• 6GB of 1600MHz DDR3 SDRAM from Corsair
• WD Caviar SE WD1600JD 160GB SATA HDD
• Tagan TG900-BZ 900W PSU
• Windows 7 Ultimate 64-bit, DirectX 11
• 30" Dell 3007WFP monitor
• NVIDIA Drivers 260.63 beta
• ATI CATALYST 10.9

• http://game.amd.com/us-en/drivers_catalyst.aspx?p=vista32/common-vista32

http://ixbtlabs.com/articles3/i3dspeed/0910/itogi-video-h2-wxp-aaa-1920-pcie.html

Puts the 5870 at 29.9 with 16x ANIS.

Updated oct 12 too. SO their newest numbers.

 

[tviceman]

According to TV's link the 5870 got 34.5 versus the leaks 34.3

Thinks to factor in:
TV's link has no AA, the leak has 4xAA
TV's link has no ANIS, the leak has trilinear

I can't find the processor but chances are it isn't a Phenom II, X4, probably a faster Core i7 with an OC.

I must have missed the part on the leak where AA is indicated in either the screen shot or in the forums. Where does the leak say it has 4xAA?