RV870 single-core, R870 dual-core, R800 quad-core?

[Helmore]

???->R300->R600->R900->???. :-)

What could RV870 be, if RV790 is 960 shaders and higher clocks @ 40nm? 1280 shaders with high clocks? Doesn't sound too big improvement over RV790.

Is it even confirmed that RV790 has 960 shaders? I wouldn't be surprised if RV770's replacement has more than that.

[Macadamia]

I'm expecting either a lot more, or none (aka speed bump).

At least 1200... for 790. ATI's shaders are small enough, and if they won't X2 this particular chip for now (maybe after shrink) they can have a lot of space to play.

[Calmatory]

Is it even confirmed that RV790 has 960 shaders? I wouldn't be surprised if RV770's replacement has more than that.

No, but that would make some sense. If the pads aren't being shrunk, then anything below the current size would make the chip pad limited. 40 nm parts are 25 % smaller, 800 -> 960 is the closest match which would increase the shaders 25%(1000). However, the shader area of the core isn't too big, so they probably could do +35% increase in shaders(1080 and beyond) whilist keeping the chip under current size.

What will be done? If it is direct increase in shaders, I wouldn't be surprised if they went +1100 for RV790 and +1500 for RV870. If they are pad limited with 256-bit MC w/ GRRD5, they can not reduce the die area too much without being pad limited again, they can only make the chip bigger. The question is, how far are they willing to go with the die size? Obviously big dies do not fit their strategy. Their superior strategy is biting them right now?

For me it seems that both cores are going to be at least as big as RV770 is. RV790 - 5-10 % bigger, RV870 - 20-33 % bigger would be my bet. Now just calculate how many shaders fit to that area and that would be my bet.

[Helmore]

No, but that would make some sense. If the pads aren't being shrunk, then anything below the current size would make the chip pad limited. 40 nm parts are 25 % smaller, 800 -> 960 is the closest match which would increase the shaders 25%(1000). However, the shader area of the core isn't too big, so they probably could do +35% increase in shaders(1080 and beyond) whilist keeping the chip under current size.

What will be done? If it is direct increase in shaders, I wouldn't be surprised if they went +1100 for RV790 and +1500 for RV870. If they are pad limited with 256-bit MC w/ GRRD5, they can not reduce the die area too much without being pad limited again, they can only make the chip bigger. The question is, how far are they willing to go with the die size? Obviously big dies do not fit their strategy. Their superior strategy is biting them right now?

For me it seems that both cores are going to be at least as big as RV770 is. RV790 - 5-10 % bigger, RV870 - 20-33 % bigger would be my bet. Now just calculate how many shaders fit to that area and that would be my bet.

How did you come up with 25% smaller? You are going from 55 nm. to 45 nm. and then to 40 nm. and I'd say you would end up about 45% smaller than at 55 nm. at least.
As for die size, they can make a smaller die by ditching Sideport and reducing power draw. The latter would mean that you would need less VDD and GND pins, which will also reduce your minimum necessary die size. So I don't think that a die size of about 200 mm^2 for RV790 would be too far off and entirely possible while still increasing unit count.

[Calmatory]

So they are not going to go directly to 40nm? Oh, you are right, ideal scaling would be 52% (40/55)^2 or 66 % (45/55)^2. So, expecting at least ~60 % or ~73 % smaller chips. Yet another sleepless night makes this thinking impossible.

So, if they are pad limited then RV790 should be at least 1280 shaders, even if they are going to 45nm first. If RV870 was 40 nm, then 1500-1600+ shaders should be very well possible.

[Helmore]

So they are not going to go directly to 40nm? Oh, you are right, ideal scaling would be 52% (40/55)^2 or 66 % (45/55)^2. So, expecting at least ~60 % or ~73 % smaller chips. Yet another sleepless night makes this thinking impossible.

No, I only meant to imply that ATI is making a full node jump. There probably will not be 45 nm. chips made by TSMC for ATI/AMD.

So, if they are pad limited then RV790 should be at least 1280 shaders, even if they are going to 45nm first. If RV870 was 40 nm, then 1500-1600+ shaders should be very well possible.

Or they will increase the capability and/or efficiency of the shaders in RV870, in which case the shaders will take up more space, but they will perform better than the actual jump in the amount of shaders would suggest. They have to add some extra logic to RV870's shaders anyway to make it compatible with DX11's specs, although I wouldn't be surprised if RV770 is already reasonably close to having all that is needed to support DX11.
RV790 will have the exact same units as used in RV770, but there will simply be more of 'em. Just my guess though, I wouldn't mind to be proven wrong.

[ColonelCain]

Or they will increase the capability and/or efficiency of the shaders in RV870, in which case the shaders will take up more space, but they will perform better than the actual jump in the amount of shaders would suggest. They have to add some extra logic to RV870's shaders anyway to make it compatible with DX11's specs, although I wouldn't be surprised if RV770 is already reasonably close to having all that is needed to support DX11.
RV790 will have the exact same units as used in RV770, but there will simply be more of 'em. Just my guess though, I wouldn't mind to be proven wrong.

I am pretty sure as well that ATI already has most of the additional logic for DX11 from their bid on DX10.1.

I also am pretty confident that at least the efficiency of the shaders will be improved, as they were able to do this in RV670 -> RV770 , which resulted in "smaller" units (more performance per mm^2 comparitively). As far as capability, that is the one that I am unsure of, as they already did some minor tweaks to the shaders to make them more GPGPU "friendly"; though who knows what changes DX11 brings to the table in the GPGPU arena (from a design standpoint).

Back to RV790 though: last time that I checked the "RV790 and RV740 in March 2009" thread, a few people were certain that RV790 would be 55nm, and just a slightly tweaked version of RV770 to yield higher clocks.
But now, you are saying that RV790 will be 40nm like RV740?

[Helmore]

I am pretty sure as well that ATI already has most of the additional logic for DX11 from their bid on DX10.1.

I also am pretty confident that at least the efficiency of the shaders will be improved, as they were able to do this in RV670 -> RV770 , which resulted in "smaller" units (more performance per mm^2 comparitively). As far as capability, that is the one that I am unsure of, as they already did some minor tweaks to the shaders to make them more GPGPU "friendly"; though who knows what changes DX11 brings to the table in the GPGPU arena (from a design standpoint).

Back to RV790 though: last time that I checked the "RV790 and RV740 in March 2009" thread, a few people were certain that RV790 would be 55nm, and just a slightly tweaked version of RV770 to yield higher clocks.
But now, you are saying that RV790 will be 40nm like RV740?

I'd say it wouldn't be wise of them to do another high(er) end chip on 55 nm. You have to consider that ATI only has a limited amount of resources and another 55 nm. would not be a wise investment of those resources. A 40 nm. chip will get them some extra practice with TSMC's new process node and it would make it easier to make a more capable chip and thus more compatible chip. A new high end chip at 55 nm. would simply not provide enough of a performance jump to be able to compete with it long enough to regain the investment you made to develop it. That's because NVIDIA will probably come in Q2 with it's 40 nm. parts, or that is probably what ATI is expecting from them. So all in all, a 40 nm. chip would be far better to do for ATI from a strategic/tactical point of view than a new 55 nm. one and that's why I'm guessing RV790 will be made on 40 nm. That's just my prediction though, although w0mbat and some other folks over at Beyond3D seem to be pointing in the same direction so to speak.

[spursindonesia]

That depends a lot on the existence of GT212, a supposed less than 300 mm^2 40 nm chip with 384 SPs 96 TMUs and 32 ROPs on 256 bot GDDR5. If that chip successfully makes appereance in late Q2, ATi MUST have something competitive against it, on 40 nm node too.

[Helmore]

ATI-Forums.de seem to think that RV790 might not even be launched at CeBit and will thus arrive later, although they aren't completely sure about that either. If that's the case than it is even more unlikely for RV790 to be made on 55 nm.
Source: http://news.ati-forum.de/index.php/n...och-kein-rv790
Google translate: http://translate.google.com/translat...90&sl=de&tl=en

[LordEC911]

ATI-Forums.de seem to think that RV790 might not even be launched at CeBit and will thus arrive later, although they aren't completely sure about that either. If that's the case than it is even more unlikely for RV790 to be made on 55 nm.
Source: http://news.ati-forum.de/index.php/n...och-kein-rv790
Google translate: http://translate.google.com/translat...90&sl=de&tl=en

Maybe because there is no RV790 since RV870 is still "relatively" on schedule.

[Calmatory]

Schedule being? Long walk till the summer, Q3 to be more precise.

[w0mbat]

Maybe because there is no RV790 since RV870 is still "relatively" on schedule.

There is a RV790, that i can tell u for shure. Recent rev is A11 w/ 750MHz coreclk and 900MHz GDDR5 memclk (just like RV770).

[jam2k]

I wonder if they're going to make a refresh of the dual gpu part ?

[ryboto]

Schedule being? Long walk till the summer, Q3 to be more precise.

being the same as it has the past few years. At least I imagine that's the implication. Process shrink after 6 months from prev. generation, then ~6 months later(1 year from previous gen. release), new generation.

[Calmatory]

being the same as it has the past few years. At least I imagine that's the implication. Process shrink after 6 months from prev. generation, then ~6 months later(1 year from previous gen. release), new generation.

That would mean RV790 is late quite a bit. Hopefully RV870 isn't.

[ryboto]

That would mean RV790 is late quite a bit. Hopefully RV870 isn't.

I used "~", plus, the 4870X2, last of the 4800 series came in Aug, 6 months from then is Feb. Also, while it's ATi's intention to do things this way, they do have to wait for TSMC to have a process ready.

[zerazax]

My edumacated guess is that a lot of the RV870 speculation might be confused with RV790 if the 750/900 clocks are true. And to provide a performance boost, it's going to have expanded units with a goal of 1.5 TFlops.

[LordEC911]

My edumacated guess is that a lot of the RV870 speculation might be confused with RV790 if the 750/900 clocks are true. And to provide a performance boost, it's going to have expanded units with a goal of 1.5 TFlops.

Or a little more. RV790 is a 40nm "fill-in" for RV870 coming out Q3/Q4?
A nice RV790 would be 1200SPs, 24SPs in 10clusters, @750mhz = 1.8TFlops.
~170mm2? 1.15b trannies? 40-50% performance increase over RV770?

Definitely leaves room for a 2000SP RV870.

[zerazax]

Or a little more. RV790 is a 40nm "fill-in" for RV870 coming out Q3/Q4?
A nice RV790 would be 1200SPs, 24SPs in 10clusters, @750mhz = 1.8TFlops.
~170mm2? 1.15b trannies? 40-50% performance increase over RV770?

Definitely leaves room for a 2000SP RV870.

Very possible too if it is 40nm

Either way, I don't believe the RV870 only has 1000-1200 SP's if its 40nm, unless they really do MCM and put two of them together. Then again, people were all over the 480 SP rumor and look what happened.

[bowman]

Or a little more. RV790 is a 40nm "fill-in" for RV870 coming out Q3/Q4?
A nice RV790 would be 1200SPs, 24SPs in 10clusters, @750mhz = 1.8TFlops.
~170mm2? 1.15b trannies? 40-50% performance increase over RV770?

Definitely leaves room for a 2000SP RV870.

1.15 billion trannies? Sounds like a crazy party.

[Helmore]

1.15 billion trannies? Sounds like a crazy party.

That's only about 200 million more than RV770 already has, doesn't sound too far off and it probably is a pretty conservative estimate. I'd say that RV790 has more than that, if it has a 256-bit memory bus.

@LordEC911: I don't think they'll change the amount of SPs per cluster/SIMD-core, that would be too much of a redesign and it would increase their ALU/TEX ratio even further (to 6:1 in your case). That's something they won't do with what is supposed to be a reasonably straight shrink. I guess we will only see an increase in the amount of clusters. I've no idea what they will do with their RBEs/ROPs, I don't think they will change the way they work and I therefor don't really expect an increase in ROPs/RBEs. Or they would add more ROP/RBE clusters with each a 32-bit memory controller instead of a 64-bit one, I think that would require the least amount of changes to the ROP/RBE section while still being able to increase the amount of ROPs/RBEs. Oh well, in the end it is all guess work for us and we may all be completely wrong.

[LordEC911]

That's only about 200 million more than RV770 already has, doesn't sound too far off and it probably is a pretty conservative estimate. I'd say that RV790 has more than that, if it has a 256-bit memory bus.

@LordEC911: I don't think they'll change the amount of SPs per cluster/SIMD-core, that would be too much of a redesign and it would increase their ALU/TEX ratio even further (to 6:1 in your case). That's something they won't do with what is supposed to be a reasonably straight shrink. I guess we will only see an increase in the amount of clusters. I've no idea what they will do with their RBEs/ROPs, I don't think they will change the way they work and I therefor don't really expect an increase in ROPs/RBEs. Or they would add more ROP/RBE clusters with each a 32-bit memory controller instead of a 64-bit one, I think that would require the least amount of changes to the ROP/RBE section while still being able to increase the amount of ROPs/RBEs. Oh well, in the end it is all guess work for us and we may all be completely wrong.

I know I had this discussion over at B3D but do you recall how exactly the TMUs are setup? Is it a quad for each SIMD? I thought that was my original assumption when looking at some of the architecture shots of R600/RV670 but I recall being told something different.

If they don't increase the number of SPs per SIMD, from 16-24, then there really isn't much they can do with RV790. They could up the SIMDs to 12 like was talked about before that isn't much of an increase.

AMD/ATi really needs to go back to back to back with strong GPUs if they really want to take over even more marketshare.

[Helmore]

I know I had this discussion over at B3D but do you recall how exactly the TMUs are setup? Is it a quad for each SIMD? I thought that was my original assumption when looking at some of the architecture shots of R600/RV670 but I recall being told something different.

The TMUs are setup differently on RV770 than on R600/RV670. On R600/RV670 a TMU block was assigned to 4 SP blocks, each block from a different SIMD 'core'. This approach has its benefits, but it is a bit tough to scale it and that's why they abandoned it for RV770. RV770 has 4 TA units and 4 TF units per SIMD 'core'. Just check here: http://www.rage3d.com/reviews/video/.../index.php?p=5
There's one odd part though, and that is that there are 'only' 32 interpolators. But that will only become a bottleneck if you INT8 bilinear filtering, you won't notice it once you apply some more AF for example and then the bottleneck will shift to the filtering units. It's still weird though, as you'd expect a texture related unit to reside with the TF/TA units and thus also scale in capability along with those.

If they don't increase the number of SPs per SIMD, from 16-24, then there really isn't much they can do with RV790. They could up the SIMDs to 12 like was talked about before that isn't much of an increase.

AMD/ATi really needs to go back to back to back with strong GPUs if they really want to take over even more marketshare.

I don't think they will increase the number of SPs per SIMD 'core', not with their first 40 nm. chips and probably not even with RV870. To me it seems like it should be fairly easy for them to increase the amount of SIMD 'cores' on a chip. They increased the amount of SIMD 'cores' that RV770 would get at the last minute when they needed to fill up some extra space, so to me it seems like you don't need to do a lot of redesigning/work to increase the SIMD 'core' count. I'd say 16 SIMD 'cores' for RV790 shouldn't be too hard to do for them.
I'd like to know what they would to with their RBEs if they do put 16 SIMD 'cores' on there though, although I'm not really sure if yet if 16 RBEs would become a bottleneck.

It will simply be guesswork for us until it's launched I guess, still interesting to think about though.

[LordEC911]

I don't think they will increase the number of SPs per SIMD 'core', not with their first 40 nm. chips and probably not even with RV870. To me it seems like it should be fairly easy for them to increase the amount of SIMD 'cores' on a chip. They increased the amount of SIMD 'cores' that RV770 would get at the last minute when they needed to fill up some extra space, so to me it seems like you don't need to do a lot of redesigning/work to increase the SIMD 'core' count. I'd say 16 SIMD 'cores' for RV790 shouldn't be too hard to do for them.
I'd like to know what they would to with their RBEs if they do put 16 SIMD 'cores' on there though, although I'm not really sure if yet if 16 RBEs would become a bottleneck.

It will simply be guesswork for us until it's launched I guess, still interesting to think about though.

Keep in mind that this won't be their first 40nm chip...
Anyways, it wouldn't be much harder to add more SPs into a SIMD than to add SIMDs. Adding more SIMDs will cause a reworking of the scheduler and will cause more overhead/latency. As for the ROPs, they sorta seem to be related to the memory controllers, like Nvidia, but not directly. I don't think more ROPs is really needed right now but maybe for RV870, ROPs are large and take up a lot of bandwidth.