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

[mczak]

Of course. VLIW had to go but I think mainly because it was poorly suited for getting into markets besides games.

GCN in a APU will be interesting because of the higher transistor count per performance level than VLIW. In the mobile market with phones and such NV is even still using non-unified architectures because of transistor count AFAIK.

I think it really isn't all that bad yes GCN loses out to some of the more efficient VLIW parts in terms of perf/transistor especially at low to medium (up to Barts) performance class (but not too badly really), but against Cayman for instance there's no such disadvantage at all (using Pitcairn of course for comparison, not the much less efficient Tahiti).
Also even at the lower end, the Mars vs. Whistler Benchmarks (HD8790M vs. 7670m) were pretty convincing - sure the former has 30% more transistors but it is also a lot more faster than that (granted if you'd scale back the clocks to the same levels it would probably be no longer spank it but still if there's any disadvantage in perf/transistor it has to be pretty small).
So GCN should rock in APUs. Especially since you can leverage for compute which at least some of the stuff you can do there isn't limited much by memory bandwidth.
(And yes it's true nvidia doesn't use unified designs for tegra, they are dead last to switch to that there, but I guess the biggest advantage it gets them is by using less accurate pixel shaders, fp20 saves a lot on datapaths, makes for way smaller multipliers etc.)

 

[mczak]

The HD 7870 Boost is a Tahiti disabled to 256-bit, and before that, the only one I can think of is the HD 2900 GT or so that was a R600 disabled to 256-bit.

Oh that one is old .
I forgot about Tahiti though which of course being a 384bit native design, isn't a power of two size neither. So if they can handle that I can't see why they couldn't handle a 192bit design.

 

[I.S.T.]

The HD 7870 Boost is a Tahiti disabled to 256-bit, and before that, the only one I can think of is the HD 2900 GT or so that was a R600 disabled to 256-bit.

I also don't think there's much of a gap between Pitcairn and Tahiti, and with the 7870 Boost there's essentially no gap. That being said, I think Hainan, which is rumored to be a double Bonaire (except for the same 2 primitives/clock), could easily replace a number of 7000 series parts. 1792 SPs at ≥1 GHz and a 256-bit bus at 6 Gbps plus a cut-down part would essentially replace both 7870s and the 7950. The 7850 still has a space but they may push 2 GB 7790 OCs to slightly close on it from below. They could even keep both Pitcairn parts and release just the full Hainan as a "7790," which would slot in in place of the 7950. So I think Hainan makes sense only when they want to replace a number of existing parts, or if they want a dual-GPU part within 300 W.

Does that include the 1 GB 7850s?

Those are being discontinued according to the Tech Report review. Apparently, the memory chips required for them are going out of production.

 

[jimbo75]

The 7850 1GB is being discontinued because of memory chips?

If AMD can convince anyone off that bullshit they deserve to sell the 7790 at the same cost.

 

[RedVi]

It uses 1Gb chips, do any other cards still use those? They couldn't use 2Gb chips without switching to a 128-bit bus, in which case the 7790 would surely outperform it.

 

[I.S.T.]

http://techreport.com/review/24539/amd-radeon-hd-7790-graphics-card-reviewed/2

 

[Kaotik]

The 7850 1GB is being discontinued because of memory chips?

If AMD can convince anyone off that bullshit they deserve to sell the 7790 at the same cost.

Same cost? There's a big difference in their price

 

[AlphaWolf]

Same cost? There's a big difference in their price

Ya, I expect it will be even bigger at retail soon enough. $150 msrp will probably get down to $135.

 

[no-X]

It's basically like, Barts was a very balanced part and Bonaire much less so.

I don't think so. Yes, TPU's results show, that Bonaire is 2-3 % slower than Barts, but the set of games seems to be a bit elderly to me. According to ComputerBase.de and Hardware.fr, HD 7790 is 28 - 29 % faster than HD 7770. Direct comparison with HD 6870 is missing, but the older reviews like this one (or another one) shows, that HD 6870 was just 18-27 % faster than HD 7770 (=> HD 7790 is very likely faster than HD 6870 in current games, maybe even quite close to HD 5870)

 

[Miksu]

Has anyone tested the 7790 against 5850? I'm wondering what's the performance difference between those two cards...

 

[Wynix]

Has anyone tested the 7790 against 5850? I'm wondering what's the performance difference between those two cards...

Differences wouldn't be very high, probably 7790 comes out in front due to it being a more efficient architecture.

Are you looking to upgrade or just making new vs old comparisons?

 

[Gipsel]

Ya, I expect it will be even bigger at retail soon enough. $150 msrp will probably get down to $135.

The XFX OC version (1075/1600 MHz) has popped up at one German online shop. It can be ordered for as less as 104.90€ including 19% VAT. That appears to be pretty cheap (would be about 115$ without VAT). It's on there for two days already, which kind of reduces the probability of an error.

 

[UniversalTruth]

Has anyone tested the 7790 against 5850

I guess the 7790 being on par (almost) with 6870, and that being considerably faster even at point of its launch, then it would be much more appropriate if somebody compares them to the 5870.

 

[Dave Baumann]

Does somebody know why AMD avoids 192 bit bus?

The initial planning configuration for this was 192b, but I was one of the proponents for moving this to 128b with the fast PHY due to the direction that the memory devices were taking. Obviously there are others trade-offs as well.

 

[Alexko]

The initial planning configuration for this was 192b, but I was one of the proponents for moving this to 128b with the fast PHY due to the direction that the memory devices were taking. Obviously there are others trade-offs as well.

Are you referring to their speed or density? Or both? I'm asking this because the Tech Report mentioned that the problem with the 1GB HD 7850 was the lack of low-density chips.

On a side note, I find it interesting that Bonaire has quite a bit more shader power than RV770/790 (14 GCN CUs vs 10 VLIW5 CUs) but less bandwidth.

It would be fun to pit them against each other, perhaps with normalised GPU clocks, and see what happens. I'm sure Bonaire would be faster, but where and by how much exactly would be interesting.

 

[Wynix]

The initial planning configuration for this was 192b, but I was one of the proponents for moving this to 128b with the fast PHY due to the direction that the memory devices were taking. Obviously there are others trade-offs as well.

It would also drive the cost up and it's already at a price point that is fairly close to the 7850, good choice making it 128b, if people need the extra memory the 7850 is what they should be getting.

I personally don't think i could buy a new card that does not overvolt anyway, i've had great fun trying to get 1000mhz core clock on my Asus 6850 using the extra voltage.

 

[Wynix]

Since you bring up density Alex, does anyone know when we will start seeing 4Gb chips used in discrete GPUs? Hynix and Samsung already have them in production so it cannnot be too far off.

 

[Dave Baumann]

Are you referring to their speed or density? Or both? I'm asking this because the Tech Report mentioned that the problem with the 1GB HD 7850 was the lack of low-density chips.

Given the comment on the fast PHY, I was primarily talking about speed. Density clearly plays a part on the planning process as well.

32Mx32 (1Gb) devices are out of production as, I believe, they are on much older process than the current or upcoming GDDR5 devices. As you may have gleaned from the PS4 announcement we are also on the cusp of another transition from 64Mx32 devices to 128M32, later on 2GB frame-buffers on 128b GDDR boards will be commonplace in this configuration and 4GB can be achieved in x16 modes.

On a side note, I find it interesting that Bonaire has quite a bit more shader power than RV770/790 (14 GCN CUs vs 10 VLIW5 CUs) but less bandwidth.

In terms of positioning / segmentation I compare 7790's ASIC to Juniper, moreso than RV770 (perhaps you can compare it to RV740), i.e. it is aimed at a slightly lower segment. But the configuration is to be expected - the longer we go into DX11 the more complex general shading becomes and more developers will lean into DirectCompute, so the more time we'll spend being shader bound rather than bandwidth bound.

 

[Davros]

whats fast PHY ?

 

[Alexko]

Given the comment on the fast PHY, I was primarily talking about speed. Density clearly plays a part on the planning process as well.

32Mx32 (1Gb) devices are out of production as, I believe, they are on much older process than the current or upcoming GDDR5 devices. As you may have gleaned from the PS4 announcement we are also on the cusp of another transition from 64Mx32 devices to 128M32, later on 2GB frame-buffers on 128b GDDR boards will be commonplace in this configuration and 4GB can be achieved in x16 modes.

Got it, thanks.

In terms of positioning / segmentation I compare 7790's ASIC to Juniper, moreso than RV770 (perhaps you can compare it to RV740), i.e. it is aimed at a slightly lower segment. But the configuration is to be expected - the longer we go into DX11 the more complex general shading becomes and more developers will lean into DirectCompute, so the more time we'll spend being shader bound rather than bandwidth bound.

Sure, it terms of positioning it's closer to chips like Juniper. I just feel a certain childish glee from comparing "old" chips with technically close(ish) new ones and marveling at the wonders of technological progress.

whats fast PHY ?

A physical memory interface optimized for high speeds.