Originally Posted by
Tempered81 @ forums.anandtech.com
Despite what charlie wrote about the new chip having next gen setup & uncore and having Evergreen shaders, it's actually rumoured to be a fairly radical change inside compared to HD2000-HD5000 series regarding 5D shader groups' composition. 1600 -> 1920, but more importantly, the 1600 Stream processors changes format from 4+1 (5 sp composition) ALU to a 3+1 '4D' or a 2 big + 2 small '4D' ALU. EG is rumoured to be a maximum of 80% efficient in its current 4+1 setup, so chopping 1 shader off wont make a big difference - but would actually increase efficiency as there would be less frequenly unused space on the core. But it will increase the chip's overall performance per square inch (or millimeter), and reduce die size.
Even without changes to the memory controller & setup engine or parts of the 'uncore' a 1600 SP core in Ati's new '4D' format would be more powerful than a 1600 SP core in the current '5D' ALU design (cypress EG). The reason is that Cypress is currently 20 SIMDs x 16 ALUs x 5 SP's = 1600sp's (320 alu). Hypothetically, with a 4D shader ALU design, a 1600sp Cypress would be arranged - for example - 20 SIMDs x 24 ALUs x 4 SP's (400 ALU total). 400 ALUs > 320 ALUs, even though both chips have 1600sp.
Starting with the HD 2900XT the ALU count basically increased to the next generation like this: 64 -> 64 -> 160 -> 320 -> 480; with the shader count increasing like this: 320 -> 320 -> 800 -> 1600 -> 1920. Since HD6870 is rumoured to have 1920SP in the 4D format, it would equal a total of 480 ALUs. 320 ALUs represents the Cypress EG core aka HD5870. The performance gain from the 4870 over the 3870 was ~92% the increase in total ALU count was 2.5 fold (64 to 160). The performance gain from the 5870 over the 4870 was ~68% and the increase in ALU count was two fold (160 to 320). This can give point to measure theoretical performance of the next gen; my personal expectation is a gain of ~40-45% over the 5870 as a result of ALU count increasing from 320 to 480. The new chip's rumored size is 40nm @ ~395mm^2 compared to the 334mm^2 Cypress and the 535mm^2 Fermi.
If they change core & memory speeds (ex: low latency 6gbps GDDR5) or much of the design of the uncore in 6870 - like the Memory controller, Rasterizers, Tessellator and Setup Engine, it would only serve to add pontential performance (shader modifications aside).
I'll be happy if the 6870 is on par with the 5970 or even a few clicks faster. This would put it 20% faster than a GTX480 and 10% faster than 512sp Fermi. Rick Bergman said Q4 2010, so my guess is Oct 15 - Nov 30th release.
Reply With Quote
Posting Permissions
