=>Jawed: I don't. I've thought about the possibility of RV740 having beefier RBEs. But do we "have definite info" that RV770 is nicely balanced in this way? Maybe the RBEs are just powerful enough that they don't limit the RV770 anywhere and if they do limit the RV740, strengthening them wouldn't improve its perf/mm2 ratio.
AMD RV770 refresh -> RV790
- Thread starter w0mbat
- Start date
It would be TWIMTBP games. I find it hard to believe that implementing DX10 reduces the performance so much on ATI hardware.
Jawed
Legend
My earlier point about RV770 "failing" with Lost Planet and Far Cry 2 appears to be due to Z fillrate - though I'm not sure. Is RV7xx unbalanced? Well, I think ATI GPUs could stand to have a higher colour:Z fillrate ratio.
There seems to be little reason to increase colour fillrate, though you can argue about blending rate.
D3D11 GPUs may make this moot, if they do not have any colour-blending and/or Z/stencil testing fixed-function units.
I still suspect RV740 has 16 RBEs with 16-bit MCs, so that each MC has a quad-RBE associated with it.
Here are some fanciful guesstimates to play with:
RV770
- 956M - total transistors
- 390M - 10 clusters (40 TUs + 800 ALU lanes + LDS)
- 566M - MCs + RBEs + L2s + shader context setup + buses + UVD + PCI Express + CrossFireX sideport
- 826M - total transistors
- 311M - 8 clusters (no double precision would make them smaller)
- 515M - MCs + RBEs + L2s + gubbins (some of which scales with 8 clusters) - CrossFireX sideport
Jawed
In this case it would be 32 RBEs/8 Q-RBEs, RV740 is 128-Bit.
I would bet on 4 ROP partitions (QRBE, MC, L2), but reduced from RV770s 64-Bit MCs to 32-Bit ones.
Jawed
Legend
Don't you just hate it when a factor of 2 goes AWOL in your thinking. Big sigh Thanks for the correction. Dunno what to think now about the mapping of RBEs and MCs.
Still puzzled why Dave apparently brought up the concept of 16-bit MCs.
Jawed
Deleted sorry.
It at least appears to be so far the other way round - they are both targetting their highest selling and minimum risk parts on the new process first. The 40nm is different to other new nodes though in the lack of competition for capacity. Stuff that was secondary previously like mobile parts and igp now has similar profitability to desktop gpus, so anything with superior volume gets pushed ahead in the line.
Although contrary to above i think i saw AMDs RS880 that was being previewed at cebit was still a 55nm chip here. I saw also nvidias mobile gpus were down 63% last quarter compared to only 34% for their desktops, so unless it really died am guessing amd has picked up some sales in mobile and hence the RV740M announcement.
Well G92 and G94 are much older than RV770 and RV730 yet AMD is going to give us "something" new before Nvidia. So in terms of turnaround time Nvidia is lagging badly regardless of what's coming in the next few months. Of course, it's debatable which strategy is better but everything about Nvidia just seems stale at the moment.
Sorry neliz, i only noticed your post of the same link in the other thread afterwards.
For the last couple of years i think amd has only ever had 3 desktop gpu's simultaneously in production at once. Even if they go well this year they probably will stay level or not quite sell as many chips as they did last year making 4 lines running at once fairly unlikely. So out of RV710, RV730, RV740, RV770 and RV790 which 2 have to go? As tchock mentioned RV710 looks to working better as a mobile part and even oems are putting in RV730 for the customers that demand more than an igp. Also if the 730 was stopped would leave quite a gap below $100 which could be awkward. Of the other chips only the RV770 is left that can be removed, they just need to be careful not to create a gap between the upmarket RV740 and base level RV790.
The price cuts they just did, coming a month before the new chips really look like stock clearing.
RV740 supposedly has UVD2 on die as well.
rjc above hit the nail on the head IMO considering the general strategy for both IHVs, when it comes to 40nm. I simply ignore any possible differences in strategy and call the whole thing crysis management.
I'd be eager to assume that anything =/>performance for 40nm has been cancelled or postponed at NV and until more reliable info appears what is going on with their =/<mainstream GT21x@40nm chips I don't think the phat lady has breathed in yet. If they're not alive then it's a totally chapter of course, but if they should be I don't expect anything from those (especially for notebook markets) than a quiet sweep under the carpet of the 40nm chips in order to further reduce manufacturing costs when they arrive.
And all of the above is still not my real point; messages from behind the curtains affecting next generation 40nm chips sound so far more than euphoristic. If you wrap the whole thing up, it looks like both have centered most of their concentration on D3D11 chips. All I'm saying here is that under the given circumstances it makes more than just sense. 3 years is more than enough chewing around on the same old architectures.
It then comes down what you really want to see under "new" exactly; you may define here the evolutionary degree of a refreshed refresh (xN) vs. a new technology generation. I may have the tendency to oversimplify things, but think over it
You guys *do* realize it's TSMC that delayed the 40nm process, right? So whether one IHV is later the the other is more related to whether they need one more respin than they thought or whether they had any last-second backend issues pre-tapeout or things like that, AFAICT.
Of course do I realize that detail, but it won't change anything in my perspective if RV870/GT3x0 tape out roughly around their projected time on 40nm.
Granted these are completely different design teams, so I'm not thinking at all about human resources. But if an IHV runs into any problems with refresh X, it should come down to the size of the (hypothetical) problem and if it's really worth in the end bothering with it or simply dump it and move on.
Silent_Buddha
Legend
Are you sure about that. I wasn't getting the impression that it was not delayed. After all it appears AMD will have 40 nm chips in Q2. When I was under the impression that the general expectation was no 40 nm chips from either IHV until Q3 or Q4.
Likewise with recent news, it appears lack of 40 nm chips for most was due to unusually low demand due to the recent economic downturn.
In other words, it appears the perceived delay isn't an unwillingness on TSMC's part, but rather an unwillingness on various companies parts to contract new chips on a new process.
Then again I have to admit I don't follow the fab industry all that closely...
Regards,
SB
waddaya know?
http://www.fudzilla.com/index.php?option=com_content&task=view&id=12439&Itemid=1
10% improvement on the 4870 would already be enough to beat the 260, but there's a bigger gap to the 285.
http://www.fudzilla.com/index.php?option=com_content&task=view&id=12439&Itemid=1
10% improvement on the 4870 would already be enough to beat the 260, but there's a bigger gap to the 285.
Jawed
Legend
5GHz GDDR5?
http://gpucafe.com/2009/03/ati-already-using-5ghz-gddr5/
25% or more bandwidth for RV790 could be useful. Erm, now with this other latest nugget about "new architecture" perhaps we are looking at a revision to the RBEs? 8x Z per clock per RBE?
Jawed
http://gpucafe.com/2009/03/ati-already-using-5ghz-gddr5/
If it's good enough for mobile it's good enough for desktop?
25% or more bandwidth for RV790 could be useful. Erm, now with this other latest nugget about "new architecture" perhaps we are looking at a revision to the RBEs? 8x Z per clock per RBE?
Jawed
Silent_Buddha
Legend
So a reworked chip and card with a new cooler. Piecing together the various rumors.
Interesting. Wonder if it reworked mostly for incremental speed increases, size, power consumption, or all of the above?
Since the sideport wasn't being used, I wonder if they either removed it entirely to save space, or fixed it to work better?
I wonder if they worked any other miracles with density and transitor packing in other areas of the chip similar to what they achieved with the ALUs previously?
Then again... Just how much were they willing to rejig things just for a (so far rumored) relatively modest speed increase? Especially with DX11 chips presumeably coming a few months after it?
Regards,
SB
Interesting. Wonder if it reworked mostly for incremental speed increases, size, power consumption, or all of the above?
Since the sideport wasn't being used, I wonder if they either removed it entirely to save space, or fixed it to work better?
I wonder if they worked any other miracles with density and transitor packing in other areas of the chip similar to what they achieved with the ALUs previously?
Then again... Just how much were they willing to rejig things just for a (so far rumored) relatively modest speed increase? Especially with DX11 chips presumeably coming a few months after it?
Regards,
SB