AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

Demo looks really smooth but it could use some AO lovin. Some parts look like they have AO, but the interface of the old mans shoes and the floor it still looks like he's floating in space.

Kudos for putting up any demo at all...we are getting less and less in the way of demos these days (games or graphics demos). Also...it might be time to start doing some OpenCL demos (physics etc). Need something to get people excited.
 
Ran pretty well on my 6950. Looked very nice. The only thing that stuck out to me compared to a CGI render was the way the old man's shoulder joints looked in motion.
 
Works on HD5 & 6 series, GF's are having troubles: http://forum.beyond3d.com/showpost.php?p=1615790&postcount=244

True that.

My 5850 Crossfire run it at 30fps average, with both cards running at 950/1250. Crossfire support is there as well although the cards are not maxing out and it is not a cpu limit. It needed the 12.2 preview to run correctly. It was crashing with 12.1.

Amazing graphics if I might add. I honestly thought that I was watching a CGI rendered video.

I wonder how well the 7970 runs it.

My 570s were showing a black screen for the most part as well! :S I wonder if Nvidia can do something to fix that, if they care.
 
lamdhd7000seriesroadmslovr.jpg


Pitcairn XT 1408SP 299$, Pro 1280 249$
Cape Verde XT 896SP 149$, Pro 832 139$

ComputerBase
 
Looks like AMD decided to clock the 7950 as low as they could without actually losing to the GTX 580. I think this suggests an attempt to maximize yields with the main constraint being the need to outperform the 580. To me, this further shows that TSMC's 28nm suffers from pretty huge variability.

I wonder how much they think they can improve upon this over time. Damien was able to take his sample down to 0.9V without touching clocks, taking perf/W to new heights. He also voiced hope that AIB partners would pick up on this and bin chips for low-voltage, releasing "Green editions" that actually deserve the name.
 
Looks like AMD decided to clock the 7950 as low as they could without actually losing to the GTX 580. I think this suggests an attempt to maximize yields with the main constraint being the need to outperform the 580. To me, this further shows that TSMC's 28nm suffers from pretty huge variability.

I think it's more to do with keeping a distance to the 7970, to make people fork even more money and buy a 7970 instead. It seems again that the scaling of SP's is quite low so further gimping was required. Until I see or hear a chip having difficulties reaching +900mhz with 1.0x voltage, I don't buy the maximizing yield argument for the chosen clock.

All in all the performance seems to be a little bit better than I expected from such specs.
 
Probably based on the questionable Lenzfire "leak".

At Newegg HD 6850 is below $150. So HD 7770 with less perfomance (see MSI roadmap), ~125mm² GPU and 128-Bit IMC should be <= $129.
The 5770 initially cost more than 4870 as well, despite somewhat lower performance.
 
The 5770 initially cost more than 4870 as well, despite somewhat lower performance.
Though it didn't take long 'till 5770 actually outperformed 4870 slightly
Probably based on the questionable Lenzfire "leak".

At Newegg HD 6850 is below $150. So HD 7770 with less perfomance (see MSI roadmap), ~125mm² GPU and 128-Bit IMC should be <= $129.
I wouldn't put too much weight on the roadmap product placement, I mean, 7950 is between 6970 and 6950 on it, too, even though it outperforms 6970 by some 15-20%, not to mention that fitting 2 new models between 6870 and 6950, when they have ~12% performance difference between them, is extremely unlikely.
 
I think it's more to do with keeping a distance to the 7970, to make people fork even more money and buy a 7970 instead. It seems again that the scaling of SP's is quite low so further gimping was required. Until I see or hear a chip having difficulties reaching +900mhz with 1.0x voltage, I don't buy the maximizing yield argument for the chosen clock.
I don't buy the yield argument neither. As far as I can tell both HD7970 and HD7950 can be overclocked to almost exactly the same, and power consumption is also similar IF the voltage is the same. But HD7950 has lower default voltage(s) which definitely eats into the overclocking potential (hence it's more like ~1000Mhz at stock voltage rather than ~1150Mhz for the HD7970), of course it also plays a role why the cards hardly draw more power than GTX560Ti and are quiet (at least some of them).

According to that roadmap, Cape Verde gets released before Pitcairn which seems a bit odd.
In any case now that we know how scaling is with more shader units, I'm still wondering what Pitcairn (and Cape Verde) looks like. The shader number (for both) somehow suggests it is possible to have CU groups of size 2 (or 1 for Cape Verde) instead of 4, or that CUs within a group can be disabled (but this seems a bit unlikely to me, since it would mean the chips in their full configuration have CU groups of dissimilar size).
 
I don't buy the yield argument neither. As far as I can tell both HD7970 and HD7950 can be overclocked to almost exactly the same, and power consumption is also similar IF the voltage is the same. But HD7950 has lower default voltage(s) which definitely eats into the overclocking potential (hence it's more like ~1000Mhz at stock voltage rather than ~1150Mhz for the HD7970), of course it also plays a role why the cards hardly draw more power than GTX560Ti and are quiet (at least some of them).

According to that roadmap, Cape Verde gets released before Pitcairn which seems a bit odd.
In any case now that we know how scaling is with more shader units, I'm still wondering what Pitcairn (and Cape Verde) looks like. The shader number (for both) somehow suggests it is possible to have CU groups of size 2 (or 1 for Cape Verde) instead of 4, or that CUs within a group can be disabled (but this seems a bit unlikely to me, since it would mean the chips in their full configuration have CU groups of dissimilar size).

http://forum.beyond3d.com/showpost.php?p=1608701&postcount=2081

Dave Baumann said:
The understanding that CU's need to be be in groups of 4, though, is a misperception.
 
Back
Top