I believe that means that texture units can read compressed MSAA buffers. Something that nvidia GPUs have been doing for a long long time.Oh, texture units reading from the compressed AA buffer for supersampling?
AMD: R8xx Speculation
- Thread starter Shtal
- Start date
I would expect that these registers are used for the group sync instructions.
Die_Hard_AMD
Newcomer
OpenGL guy
Veteran
That wouldn't give you supersampling.Oh, texture units reading from the compressed AA buffer for supersampling?
I don't see any connection, ssaa != msaa.
Yeah hence my confusion (I got that from the Rage3D link btw)
Silent_Buddha
Legend
Hmmm, has there always been a "CrossfireX Compositor" on ATI cards? I can't remember seeing that on Rv770 layouts.
Regards,
SB
Regards,
SB
At first glance I thought the diagram said 'CrossfireX Connector'. Either way would having the CFx unit directly on the hub allow for shared RAM finally? Or at least quicker access to each others RAM?
Looks like ATi has been beavering away on CF quite a bit this time around. Maybe this is the working version of what was supposed to be on the 48xx?
-Plack
This makes me a saaaad panda
I was really gunning for some way to either split a single cypress into two junipers, or an easier way to glue two cypress' together
-Plack
GPUs like GT200 or RV770 can call them heterogeneous multi-core designs and GPUs like SGX543MP can call them homogeneous multi-core designs.
RV870 from the early details that we have is kinda similar design with RV770.
For architectures like RV870 MCM designs makes no sense at all imo. (regarding performance scaling / die scaling ratio...)
tweakers.net is doing a very good job...
Lots of info inside...
I wonder who lltech4 is...
Regarding the AMD benchmarks (if they are real...) i think that for 4X AA 16X AF DX10.0 based codepaths, a good performance indication is the following:
5850 will be from -10% up to +35% in relation with GTX285
5870 will be from +10% up to +60% in relation with GTX285
For some games implementing DX10.1 codepaths probably 58XX will be faster than that...
So I guess 58XX is bandwidth limited.
I suppose the 5870 would need something like 1,5GHz instead of 1,2GHz to show its true colors and 5850 something like 1,2GHz.
AMD is making a hell of a pre launch marketing job...
RV870 from the early details that we have is kinda similar design with RV770.
For architectures like RV870 MCM designs makes no sense at all imo. (regarding performance scaling / die scaling ratio...)
tweakers.net is doing a very good job...
Lots of info inside...
I wonder who lltech4 is...
Regarding the AMD benchmarks (if they are real...) i think that for 4X AA 16X AF DX10.0 based codepaths, a good performance indication is the following:
5850 will be from -10% up to +35% in relation with GTX285
5870 will be from +10% up to +60% in relation with GTX285
For some games implementing DX10.1 codepaths probably 58XX will be faster than that...
So I guess 58XX is bandwidth limited.
I suppose the 5870 would need something like 1,5GHz instead of 1,2GHz to show its true colors and 5850 something like 1,2GHz.
AMD is making a hell of a pre launch marketing job...
Well.. not always. I fondly remember the good old days of the Crossfire Mastacard Edition.
All you need to get CrossFired up!
Agreed. I was wondering why they did a few things the way they did, but after looking around at all the interest they've been generating I think they know what they're doing.
Q2/08: -$121M (lost)
Q3/08: +$62M
Q4/08: -$148M (lost)
Q1/09: -$201M (lost)
Q2/09: -$105M (lost)
total: -$513M for last 5 quarters
average: -$103M per quarter
As I said, -$12M for ATi's last quarter is nothing compared to these numbers.
Jawed
Legend
Dude, what's your point? AMD financials:
Q2/08 -$1,400M (lost)
Q3/08 -$400M (lost)
Q4/08 -$1,420M (lost)
Q1/09 -$414M (lost)
Q2/09 -$330M (lost)
That's disregarding the 6 losing quarters before that, mind you. Slappi's assessment was a response to someone claiming ATI was profitable. In fact, it just lost less than the rest of the company.
Back to the hardware:
One thing that has some ho-hum scaling is the internal bandwidth.
The L1 numbers seem to be in line with what would be expected with the doubling of the SIMD count.
The L1/L2 bandwidth appears to be in line with clock speed, even if the consumers of said bandwidth are twice as numerous.
This may be one area where some of the less than doubled performance might be attributed.
One thing that has some ho-hum scaling is the internal bandwidth.
The L1 numbers seem to be in line with what would be expected with the doubling of the SIMD count.
The L1/L2 bandwidth appears to be in line with clock speed, even if the consumers of said bandwidth are twice as numerous.
This may be one area where some of the less than doubled performance might be attributed.
What about external bandwidth? Any confirmation yet? That kinda of scaling would be amazing if they were still on a 256-bit bus.
By the performance or architectural changes (or lack thereof)?
By the performance or architectural changes (or lack thereof)?
Just got a PR thingy from AMD, thought it might fit in here somewhere. Story in full below so no one can whine that I'm hit whoring :
:digitalwanderer at Elite Bastards said:
Similar threads
