Is this accounting for the effect of the L2 cache?
AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread
Imagination claims to have lossless color compression with a 2:1 average ratio. Maybe Nvidia does too. It's just the first thing that came to mind other than caching effects.
DavidGraham
Veteran
I like to think performance is affected too. Theoretically, a highly clocked, interface will not be able to sustain peak bandwidth all the time. DDR tend to cut bandwidth in half for address data and DDR5 needs some training to run at max frequency .
So depending on the frequency, a 512-bit bus could always be better than a 384-bit bus or vice-versa.
No. AMD would profit too, at least a bit. It's probably just what I said, a bunch of screen filling quads.I think this is a proprietary benchmark they are testing with and it could be using optimized data set to align/fit the screen tiles nicely into the cache?
Is nV using the L2 also for caching ROP acesses? Anyway. It should trash the caches by design of the benchmark as the render target is much larger than the L2 and every pixel is only written once before the next quad comes. That's how I came to the idea above.
They have a TBDR. It's probably much easier to implement lossless compression for a tile there. For an IMR one wants relatively small tiles to load and store for ROP accesses to reduce the bandwidth overhead. But this makes general compression techniques harder to implement I guess (and the reason why compression is used only for MSAA render targets and bases on the simple fact that multiple samples belonging to the same triangle have necessarily the same color). Maybe hardware.fr should make fillrate tests with randomized content
.Fermi made the L2 service all memory clients.
http://www.beyond3d.com/content/reviews/55/9
It is an assumption of mine that the arrangement is unchanged for it successor.
The 512kB of a GK104 are too small in any case, if not a second thing plays into this.
So we are back to bandwith constrain there.. if your bandwith dont follow, the pixel fill rate advantage become useless. ( i refer to the high res / high AA level performance difference between Kepler and Thaiti ) .
Hence why maybe you see the advantage on "pure" pixel fillrate benchmark ( 3Dmark etc ) and not on real usage . Look like AMD just choose a " right balance ". Increasing a little bit the number of Rop, L2 cache size and a bit more bandwith ( same 384bits bus + higher memory speed ), should provide a right config ( without having to over specs one or the other ) .
Last edited by a moderator:
since 4K gaming has come up as an issue with tearing and on current eyefinity the tearing on current cards and drivers is the "fix" from amd to make the new generation cards for 4K gaming the solution and not the older gen?
One reason I havent gone crossfire with my 7970 as the issues with eyefinity have a tendency to be problematic to say the least when things dont work as advertised.
http://techreport.com/blog/25399/here-why-the-crossfire-eyefinity-4k-story-matters
One reason I havent gone crossfire with my 7970 as the issues with eyefinity have a tendency to be problematic to say the least when things dont work as advertised.
http://techreport.com/blog/25399/here-why-the-crossfire-eyefinity-4k-story-matters
I think a large part of the problem is that they tried pushing 4K too early; current cards just don't cut it.
A recent analysis of Ultra High Definition Gaming by Guru3d came to the same conclusion regarding crossfire/eyefinity.
http://www.guru3d.com/articles_pages/ultra_high_definition_pc_gaming_benchmark_review_uhd,12.html
http://videocardz.com/45704/amd-radeon-r9-290x-hawaii-gpu-pictured-512-bit-4gb-memory
PCB shots from DICE, confirming apparently 512bit membus and rectangular die rather than square
DICE also leaked 7990, didn't they?
PCB shots from DICE, confirming apparently 512bit membus and rectangular die rather than square
DICE also leaked 7990, didn't they?
Why is this not in its proper thread?
Looks to be close to that 425mm² estimate we had already from the "GK110 is 30% larger" quote.
Or they have large orders/contracts to fill and/or the PRO version will compete with Nvidia's offerings.
Tahiti made use of a crossbar to decouple the ROPs from the memory controllers.
I would say 48ROPs is much more likely than 64ROPs.
Very true. They did say they are aiming at the enthusiast segment (id take that as 7950/7970 territory) rather then "ultra enthusiast" (7990?). Makes me think they may have had a full die top tier consumer planned before the 780, realized they could get away with a cut down and meet performance comparability and pushed off launch to fill stock.
This also leaves them with a comeback to a titan "ultra" full die, or even a middle card for between VI and PI