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

Am I the only one who thinks this price discussion belongs more in 3D Hardware, Software & Output Devices than here, as price is something that relates to actual products and not architectures/chips?

I think we should stick to architectural discussion in this section.
 
I wonder what other consumer realm you see a company offering less, asking for more and have consumers on forums thinking it's great.

That's pretty much been the case with most of Nvidia's cards for the past 2 years. Best example would be the 560 Ti which for a long time cost more than the faster 6950 yet is still lauded as the greatest card ever in many quarters.
 
Why are people solely focusing on the consumer retail space? The real volume (tens of thousands of units) consumers of this product will be OEMs. You are also missing the fact that OEMs negotiate chip supply contracts months in advance with prices locked in at the moment of purchase, which makes falling prices at retail (of EOL'd products by the way) largely meaningless to them.

Apple is still selling the 5770 and 5870 in their Mac Pro. At what price do you think they bought all those chips? I'm fairly certain they paid at least as much if not more than what will cost them to upgrade to the new AMD products(don't forget about volume discounts). Hell the HD7770 launched at the same exact price the 5770 launched.

I would not be surprised to see Cape Verde all over OEM boxes, because the power consumption is simply outstanding which allows them to upgrade the system in other areas and as an added bonus marketing can use the new name and larger bulletpoint list of features to their advantage.

P.S: Apple is also still selling the 6970m (Barts Pro based) in their iMac and 6770m and 6750m (Juniper based) in their MacBook Pro.
 
Last edited by a moderator:
P.S: Apple is also still selling the 6970m (Barts Pro based) in their iMac and 6770m and 6750m (Juniper based) in their MacBook Pro.
6750m/6770m are not Juniper based (Turks instead).
I agree though Cape Verde would make an excellent notebook part. Probably with slightly slower gddr5 (really need those 1.35V chips), clocked slightly slower than the 7750 but with all CUs enabled it would easily beat the fastest Juniper-based mobile chips while drawing less power. Sort of strange those mobility chips were hardly ever mentioned in any credible rumors.
ZeroCore is actually sort of a waste in notebooks since you can completely disable the chips via acpi methods (which saves that last watt or so too). Though it would probably help a lot with the considerable amount of OEMs not being able to get any kind of bios table right :).
(Hmm actually I might be wrong about this. Reading anand's article about optimus vs. dynamic switchable AMD graphics again, it seems like dynamic switchable always worked like that - i.e. ZeroCore isn't actually new on Southern Island just didn't have a catchy name before and wasn't enabled on the desktop parts. Not sure though maybe notebooks still do more, in any case if it helps getting rid of OEM custom drivers which are usually totally useless and you need hacks to install something sensible all the better.)
 
Last edited by a moderator:
After reading the reviews, I feel that AMD probably made a small mistake in going with 10CUs. I think that the primary reason was to get the CapeVerde Pro to be without the 6-pin, highly doubt that adding 2 more CUs would have increased the die size by a significant margin. Looking at the power consumption figures, they could have gone with 12CUs (12 for XT, 10 for Pro) and still could have done without the 6-pin on the Pro model. In such a hypothetical configuration, the XT would have been close to the 6870/560 and Pro would have been close to the 6850.

Again, its all about forecasting and I doubt they would have been comfortable with going with 12CUs if that (6-pin less) was really the internal target guideline.
 
After reading the reviews, I feel that AMD probably made a small mistake in going with 10CUs. I think that the primary reason was to get the CapeVerde Pro to be without the 6-pin, highly doubt that adding 2 more CUs would have increased the die size by a significant margin. Looking at the power consumption figures, they could have gone with 12CUs (12 for XT, 10 for Pro) and still could have done without the 6-pin on the Pro model. In such a hypothetical configuration, the XT would have been close to the 6870/560 and Pro would have been close to the 6850.
No that wouldn't have been enough for that. I have to grossly guess here as far as scaling with CUs is concerned as I haven't seen a single review comparing 7770 and 7750 at the same clock (with PT+20% of course if overclocking the 7750), but I think a somewhat fair assumption would be for 20% more CUs you still couldn't expect more than a 10% increase in performance (judging by separate core/mem overclocks the memory actually seems to start limiting quite a bit if you only increase core clocks), and I think that's quite optimistic already. So with the reference clocks 12 CUs at best would have helped to catch the 6850. Likewise the Pro (which should see more of an increase because not only the difference percentage-wise is larger but also the shader/bandwidth ratio is significantly lower) would probably go from a bit slower than 5770 to a bit faster on average (and very obviously nowhere near 6850 since the current 10 CU part with much higher clocks can't even get the 6850).
Again, its all about forecasting and I doubt they would have been comfortable with going with 12CUs if that (6-pin less) was really the internal target guideline.
I don't know if it was due to this, AMD could have fused off more CUs to still hit that target with the pro (why not have 9 CUs...), though admittedly that would only have increased the already large difference between the 7770 and 7750 (and hey I criticized some of the parts as basically non-scaling with simd count in the past so I shouldn't complain too much now that AMD "fixed" that "problem" by lowering the CU count already on the 7750, though scaling _might_ be better now). By the die shots and some pixel counting a CU is ~5mm² (though that does not include any area potentially needed outside the visible regular structure) so a 2 CU increase, while not that dramatic, definitely makes a difference with such a small chip.
 
No that wouldn't have been enough for that. I have to grossly guess here as far as scaling with CUs is concerned as I haven't seen a single review comparing 7770 and 7750 at the same clock (with PT+20% of course if overclocking the 7750), but I think a somewhat fair assumption would be for 20% more CUs you still couldn't expect more than a 10% increase in performance (judging by separate core/mem overclocks the memory actually seems to start limiting quite a bit if you only increase core clocks), and I think that's quite optimistic already. So with the reference clocks 12 CUs at best would have helped to catch the 6850. Likewise the Pro (which should see more of an increase because not only the difference percentage-wise is larger but also the shader/bandwidth ratio is significantly lower) would probably go from a bit slower than 5770 to a bit faster on average (and very obviously nowhere near 6850 since the current 10 CU part with much higher clocks can't even get the 6850).
12CUs at 1GHz would have a higher shading power than the 6850. In it's current form the 7770 is anywhere from 5%-11% slower (depending on the review) than the 6850 while having 16% less shading power. Basing on that and with some driver optimization, the theoritical 12CU part could have ended between the 6850/6870.

I don't know if it was due to this, AMD could have fused off more CUs to still hit that target with the pro (why not have 9 CUs...), though admittedly that would only have increased the already large difference between the 7770 and 7750 (and hey I criticized some of the parts as basically non-scaling with simd count in the past so I shouldn't complain too much now that AMD "fixed" that "problem" by lowering the CU count already on the 7750, though scaling _might_ be better now). By the die shots and some pixel counting a CU is ~5mm² (though that does not include any area potentially needed outside the visible regular structure) so a 2 CU increase, while not that dramatic, definitely makes a difference with such a small chip.
It would still have been a smaller chip than Juniper.
 
Well I just scaled up performance by a factor of 2 (which is about GTX 580) then subtracted a little bit to account for not perfect scaling :). Though you're probably right that would be a bit too good scaling...

Still AMD can mitigate that problem by scaling bandwidth above 100%. I'm expecting either 5GHz or 5.5GHz 256bit memory on Pitcairn.
 
Still AMD can mitigate that problem by scaling bandwidth above 100%. I'm expecting either 5GHz or 5.5GHz 256bit memory on Pitcairn.
Could be to much for the sweet-spot ~250-260mm².
After looking on 512KiB L2 in CV (Tahiti 768), we could see 1MiB in Pitcairn, which should uplift the perfomance with only 4,5Gbps GDDR5.
 
After reading the reviews, I feel that AMD probably made a small mistake in going with 10CUs. I think that the primary reason was to get the CapeVerde Pro to be without the 6-pin, highly doubt that adding 2 more CUs would have increased the die size by a significant margin. Looking at the power consumption figures, they could have gone with 12CUs (12 for XT, 10 for Pro) and still could have done without the 6-pin on the Pro model. In such a hypothetical configuration, the XT would have been close to the 6870/560 and Pro would have been close to the 6850.

Again, its all about forecasting and I doubt they would have been comfortable with going with 12CUs if that (6-pin less) was really the internal target guideline.

I posted the exact same thing a few pages back :smile:

http://forum.beyond3d.com/showpost.php?p=1620427&postcount=2750

Agree, the price/performance ratio of Cape Verde is terrible. Its actually a Turks replacement which AMD is trying to push on us as a Juniper/Barts replacement..and at Juniper/Barts prices no less :rolleyes: At least HD 7970 and 7950 beat their competitors convincingly, and were priced either the same or lower. Here we have HD 7770 barely reaching HD 6850/GTX 460 while launching at a good $10-20 higher.

I really wonder why they didnt go for a 12 CU part instead of 10. Adding 2 more CU's would have added very little to the die size, say 10 mm2. Surely the cost difference between a 130 mm2 chip and a 120 mm2 chip cant be that high. And the memory clock is on the lower side as well. Maybe they reused the old Barts MC because of the smaller size. With 20% more CU's and BW (say if it was clocked at 5.5 Ghz), the performance might have still been acceptable, and would have pushed it into the next segment. The cost increase would have been negligible.

I was actually looking forward to Pitcairn..but now i dont think i'll keep my hopes up :cry:

Edit: Also for a 12 CU part, if they launched a cut down part with 2 CU's disabled, the differnce would have been lower(16%) than between a 10 CU and 8 CU part (20%)
12CUs at 1GHz would have a higher shading power than the 6850. In it's current form the 7770 is anywhere from 5%-11% slower (depending on the review) than the 6850 while having 16% less shading power. Basing on that and with some driver optimization, the theoritical 12CU part could have ended between the 6850/6870.

It would still have been a smaller chip than Juniper.

It would have been far smaller than Juniper. Juniper is 166 mm2. By my estimation it would have been around 135 mm2

Could be to much for the sweet-spot ~250-260mm².
After looking on 512KiB L2 in CV (Tahiti 768), we could see 1MiB in Pitcairn, which should uplift the perfomance with only 4,5Gbps GDDR5.

I doubt they'd go for 1 MiB in the mid range chip. I think we'll see the same 512 KiB as Cape Verde.

Still AMD can mitigate that problem by scaling bandwidth above 100%. I'm expecting either 5GHz or 5.5GHz 256bit memory on Pitcairn.

If Pitcairn is really only 2X Cape Verde i.e 20 CU's, it may just carry forward the same memory controller and hence the same slower 4.5 Ghz speed. But if its anything more than 20 CU's, i hope they go for the higher speed 5/5.5 Ghz controller.
 
I think some peoples thinking is very desktop centric. Obviously there are more parameters at play during product planning.
 
So, what are the 7750 and 7770 going to be called in laptops?

As I suggested earlier I imagine the 7750 could even fit a laptops TDP needs at its desktop speed. The 7770 should be near the 6990m in performance, but could it be a 1Ghz laptop part?
 
Since I can't see it having been done yet, let me congratulate those at AMD who sometimes visit these boards on having gotten their second 28nm graphics chip out the door!
Helped by process and design both, it seems to deliver both excellent power characteristics generally, and performance/watt specifically which will benefit all its users, together with an impressive and wide ranging feature set.

Kudos.
 
It would have been far smaller than Juniper. Juniper is 166 mm2. By my estimation it would have been around 135 mm2.
Yes but that's still a lot larger than Turks.
Plus I have really no idea why people are thinking 12 CUs would get it up to 6870 level. Even a 20% core overclock only gets you about a 10% performance increase (with constant memory clock, as seen here: http://www.hardware.fr/articles/855-20/gains-avec-overclocking-gpu-memoire.html) so a 20% increase in CUs almost certainly would be less than that. I don't disagree that 2 more CUs would have been nice, but it wouldn't have changed things much.
Of course with 2 more CUs, higher core AND higher mem clock it could get somewhere between 6850 and 6870 level. But for mobile the chip couldn't be pushed anywhere close to that anyway.

If Pitcairn is really only 2X Cape Verde i.e 20 CU's, it may just carry forward the same memory controller and hence the same slower 4.5 Ghz speed. But if its anything more than 20 CU's, i hope they go for the higher speed 5/5.5 Ghz controller.
I'm not sure it's really a low-speed gddr5 design. Overclocking seems to indicate otherwise. Plus the fact that I've yet to see any amd gpu where not either the memory is overvolted or the clock is significantly lower than the gddr5 chip maximum (and here it's simply the latter), so imho the somewhat "low" memory clock speed does not indicate the controller wasn't designed to handle more (probably not 6Ghz but 5Ghz seems quite possible). Plus for the faster chips it probably makes more sense to go for the more expensive gddr5 chips anyway.

So, what are the 7750 and 7770 going to be called in laptops?
Educated guessing suggests 78xx? Though I think the specs will be a bit lower as usual, the top dog might not exceed desktop 7750 performance by much.
 
I think some peoples thinking is very desktop centric. Obviously there are more parameters at play during product planning.

Eheh, I've been thinking, ever since I saw the reviews for Cape Verde: this really looks like a laptop GPU.

It's small, cheap to manufacture, but still much much faster than anything integrated, even in Trinity. It can be a great fit for medium-sized gaming-oriented laptops, but also find its way into cheaper, smaller models. Factor in PowerTune to let OEMs pick their TDP, throw in ZeroCore for battery life, and you've got yourself a winner.

Basically, I see Cape Verde as the perfect chip for the mobile GPU sweet-spot, able to target most mainstream notebooks.
 
Educated guessing suggests 78xx? Though I think the specs will be a bit lower as usual, the top dog might not exceed desktop 7750 performance by much.

Well the current top dog is essentially a 715mhz 6870 (6990m), I expect there will be a 256 bit bus top dog again this gen, a pitcairn derivative presumably would be a decent jump over the 7750 which I don't think threatens the 6990m.
 
I posted the exact same thing a few pages back :smile:

http://forum.beyond3d.com/showpost.php?p=1620427&postcount=2750
Good to see I'm not the only one here that thinks this way.
I think some peoples thinking is very desktop centric. Obviously there are more parameters at play during product planning.
So I guess thats an indirect confirmation that TDP was the biggest factor :p (I never meant that us forum dwellers have more insight than anyone from your group. :cool:)
 
Well the current top dog is essentially a 715mhz 6870 (6990m), I expect there will be a 256 bit bus top dog again this gen, a pitcairn derivative presumably would be a decent jump over the 7750 which I don't think threatens the 6990m.
Yes agree. Really meant "top Cape Verde chip" there. Should definitely see some pitcairn in notebooks, but of course restricted to the big/heavy category.
 
So, Juniper was a little too big for the notebook segment it was in then? Those chips do run quite hot in notebooks, although that's usually because of a barely adequate cooling design sometimes even in big notebooks (ie G73JH at >100C).
 
Back
Top