AMD Radeon 7000-Series 28nm (Southern Islands) | 7990 7970 7870 7770 | Discussion

[RussianSensation]

Just sold my GTX 580. Luckily I had some accumulated Discover reward points, so I used them to pick up an HD 5450 to pop in here 'till the 7970 is out!

Smart. GTX580 is peaking in price right now since it's the holiday season and will prob. drop in price once 7970 is out. How is gaming on the 5450?

 

[Skurge]

Anyone thinking if Tahiti has a 384bit bus, it's going to have a larger die than Cayman? Cause at a certain die size, getting a 384bit bus doesn't cost much as the die is so large there is enough space for it.

So it could end up being 28nm AND have a larger die than Cayman. I can see it being very fast. 60%+ faster than Cayman and that would be before you add in higher clocks. If it really comes out at 1GHz.

 

[Pacman4]

How so? A video popped up on top of the image, one which I couldn't close. And then my symantec antivirus said trojan detected?

It's just some pop up box asking you to fix some vid issue, it may or may not be suspect, but don't click on it, close it as I did.

 

[Arachnotronic]

Smart. GTX580 is peaking in price right now since it's the holiday season and will prob. drop in price once 7970 is out. How is gaming on the 5450?

Hehe, it's final exam week -- no time for gaming!

At home I'll be building myself a budget little Llano rig for basic gaming until 7970 hits the street. Hopefully the money I got from my GTX 580 will appreciate in the stock market before I pull it out to buy the 7970 xD

 

[Silverforce11]

Anyone thinking if Tahiti has a 384bit bus, it's going to have a larger die than Cayman? Cause at a certain die size, getting a 384bit bus doesn't cost much as the die is so large there is enough space for it.

So it could end up being 28nm AND have a larger die than Cayman. I can see it being very fast. 60%+ faster than Cayman and that would be before you add in higher clocks. If it really comes out at 1GHz.

GCN is still an unknown in terms of perf/mm2 and perf/watt.

I think we won't know or our speculation (based on specs) will not be anywhere near correct until some proper benchmark leaks.

 

[Arachnotronic]

The shrink to 28nm alone pretty much guarantees a significantly faster GPU.

 

[MrK6]

The shrink to 28nm alone pretty much guarantees a significantly faster GPU.

Was Prescott that long ago?

 

[RussianSensation]

Anyone know why AMD doesn't add eDRAM to modern discrete GPUs as it was done for Xbox360? I thought that gives you almost free 4x AA? How come this design isn't used on modern GPUs?

 

[boxleitnerb]

Bandwidth is so large nowadays, I don't think it would make that much of a difference.

 

[ocre]

Anyone know why AMD doesn't add eDRAM to modern discrete GPUs as it was done for Xbox360? I thought that gives you almost free 4x AA? How come this design isn't used on modern GPUs?

Wow, there is the 7000 series!!!! 64mb DDR! Must be a beast!

 

[Blitzvogel]

eDRAM is costly (it's not only memory, it's a more complicated circuit board too) and DX I don't think has any provisions for it. It probably increases the power draw and heat by a large enough amount too.

 

[tviceman]

Anyone thinking if Tahiti has a 384bit bus, it's going to have a larger die than Cayman? Cause at a certain die size, getting a 384bit bus doesn't cost much as the die is so large there is enough space for it.

So it could end up being 28nm AND have a larger die than Cayman. I can see it being very fast. 60%+ faster than Cayman and that would be before you add in higher clocks. If it really comes out at 1GHz.

I have no idea where the die size will end up, but if the 384-bit rumors turn out true then I think that is really good news all around.

 

[BFG10K]

I want that evil wizard card! Now! :biggrin:

 

[wlee15]

Anyone thinking if Tahiti has a 384bit bus, it's going to have a larger die than Cayman? Cause at a certain die size, getting a 384bit bus doesn't cost much as the die is so large there is enough space for it.

So it could end up being 28nm AND have a larger die than Cayman. I can see it being very fast. 60%+ faster than Cayman and that would be before you add in higher clocks. If it really comes out at 1GHz.

Remember that Cayman had a fairly large die itself with an area of 389 mm2. R600 after all was 420 mm2 with a 512-bit bus.

 

[Arkadrel]

Remember that Cayman had a fairly large die itself with an area of 389 mm2. R600 after all was 420 mm2 with a 512-bit bus.

Cayman was pretty fail.... was way to big vs its performance.
If you look at the 6870 @ 255mm^2, its pretty close to the 6950 in performance, but on a much smaller die.

I hope they do better with the 7xxx series than they did with cayman.

 

[Grooveriding]

Where does it say that? > can mean 1% or it can mean 100%



http://forum.hardware.fr/hfr/Hardware/2D-3D/unique-southern-islands-sujet_877763_117.htm#t8143200

6+8pin connected, but on the card there are 2x8pin connectors

12 memory chips -> 384 bit interface

2x8Pin for a stock single gpu card would be over the top. Nice, but over the top :biggrin:

 

[Lonbjerg]

Cayman was pretty fail.... was way to big vs its performance.
If you look at the 6870 @ 255mm^2, its pretty close to the 6950 in performance, but on a much smaller die.

I hope they do better with the 7xxx series than they did with cayman.

I guess some people havn't figured out what GNC mean for the architecture.

AMD's VLIW4(and VLIW5) was well suited for games.
But for GPGPU it was either great in a few instances...or abysmal in most places.
And GPGPU is not going anywhere, quite the opposite...from PhysX to Vreveal.
AMD needed a new architechture...or loose the battle.

What does this mean?
it mean they will oose gaming performance per "core" with GNC comapred to VLIW4...but get out of the "mostly bad, sometime great GPGPU" performance they have now.

Please have this in mind before guessing on perfomance...otherwise you will end up with a "Dulldozer" feeling...

 

[tviceman]

What does this mean?
it mean they will oose gaming performance per "core" with GNC comapred to VLIW4...but get out of the "mostly bad, sometime great GPGPU" performance they have now.

Please have this in mind before guessing on perfomance...otherwise you will end up with a "Dulldozer" feeling...

This is exactly what is going to happen. There is no doubt Tahiti will be better than cayman in terms of raw performance and perf/watt, but the whole point of creating GCN was, for a lack of better words, to be more like Nvidia's flagship chips - serve multiple products for different markets and serve them all well. When comparing GCN against VLIW4, concessions will be made.

 

[Arachnotronic]

Hell, even if GCN is a 28nm Fermi, it'll still blow the doors off of a 40nm Fermi.

 

[Lonbjerg]

Hell, even if GCN is a 28nm Fermi, it'll still blow the doors off of a 40nm Fermi.

Good for AMD that NVIDIA stopped making GPU's after "Fermi" then...oh wait :whiste:

 

[Arachnotronic]

NVIDIA's high end lineup is nowhere in sight. AMD's is right 'round the corner.

 

[Lonbjerg]

NVIDIA's high end lineup is nowhere in sight. AMD's is right 'round the corner.

And that matters why?
Both the perfomrnace of AMD's and NVIDIA's next generation is unkown.
But you go right ahead and preorder.

I'll await to see the benches...when both generations are out.
I have no "faith" in AMD pre-hype...GNC will up GPGPU in favour of gaming...NVIDIA already crossed that brigde with the G80.

 

[iCyborg]

I don't care for these price points. I guess we will have to see what the actual benchmarks show, but I was hoping the 7950 would hit the 349 area not push 400.

I have mixed feelings about 7950 costing $400-$450:
- on one hand it should indicate great performance which is good
- on the other hand, $450 is generally higher than I usually shell out for a GPU

 

[iCyborg]

I'll await to see the benches...when both generations are out.
I have no "faith" in AMD pre-hype...GNC will up GPGPU in favour of gaming...NVIDIA already crossed that brigde with the G80.

That's why I'm not buying IB next spring/summer. I'm gonna wait until 2014 when AMD will also have ~22nm and then decide

Yeah, G80 upped the GPGPU in favor of gaming, we all remember how 8800GTX was lackluster in terms of gaming...

 

[RussianSensation]

Yeah, G80 upped the GPGPU in favor of gaming, we all remember how 8800GTX was lackluster in terms of gaming...

I think you misinterpreted his comment. What he meant was that NV's performance/watt suffered significantly after they scaled the GPU architecture to be more suitable for both the gaming and GPGPU markets. But now that they are over that hurdle, they can just focus on scaling the Fermi architecture and improving efficiency further since the transition to a brand new advanced GPGPU architecture has occurred.

On the other hand, AMD is only now going multi-purpose GPU route and we don't know how that will actually translate into performance/watt per transistor. It's not unreasonable to assume that what happened to NV is going to happen to AMD too (i.e., reduction in performance/transistor).

Let's say HD7970 could have had 2560 SPs if it simply went to 28nm and reused VLIW-4. It could be that because you now have 30-32 CUs, all that additional GPGPU complexity, you might only be able to fit 2048 SPs into the same die space. At the same time, it's not unreasonable to think that there is a chance that AMD's scalar architecture is more efficient than Fermi's.

Obviously, this is just random speculation on my part. Even once HD7970 launches, it will be difficult to assess how good it really is without having NV's 28nm GPU to compare (unless of course we get 75-100% more performance over 6970).

Still, we must not underestimate the die shrink. The 55nm HD 4870 packed 956m transistors into a die size of 263mm² while the 40nm HD 5870 had 2.15bn transistors - 2.25 times more - in a die area of 334mm², which is only 27% bigger than the previous one.

Even if on a performance/transistor basis GCN is less efficient than VLIW-4, in terms of performance/watt (not per transistor), HD7970 can still be superior to the HD6970 simply because the shrink from 40nm to 28nm is so significant that they'll be able to fit more SP, TMU and ROP units regardless that some extra die space will be allocated for GPGPU aspects.