AMD: R9xx Speculation

[Broken Hope]

Looks like default is the lower TDP setting, or the newer drivers used here are a lot better.

I can confirm the German results are fake.
Well, my 6970 from Sapphire arrived.
They've specifically gave a download exclusively to the latest drivers which truly unlock the performance, and it's beating a stock 580 in Heaven, albeit the max FPS is a little lower.

 

[mczak]

3072 Steamprocessors

I haven't seen anything yet which said it's 2x full cayman. I guess though with the new dynamic clocking scheme AMD got a lot more room to play with clocks instead of disabling simds for power reasons, so it might make sense.

 

[Bouncing Zabaglione Bros.]

More from Gibbo at OCUK:

Hi there

Some OC's from my testing:-

HIS 6950 stable at 930Mhz Core and 5950MHz Memory.

HIS 6970 stable at 980MHz Core and 6200MHz Memory.


I think there is more to go with voltage tweaking etc.

Has AMD pulled off a double bluff, and 6970 is going to easily match or better GTX580 with a significantly smaller and cheaper chip? With Antilles waiting to do more damage next month?

 

[Mize]

Has AMD pulled off a double bluff, and 5970 is going to easily match or better GTX580 with a significantly smaller and cheaper chip? With Antilles waiting to do more damage next month?

(Assume you meant 6970)
While I seriously doubt the 6970 will beat the 580, it would be wonderful for all of us (fans of either or none) if it did...lower prices!!

 

[trinibwoy]

No shit? That would be quite a coup after all the flak they got over tessellation.

 

[Razor1]

More from Gibbo at OCUK:

Has AMD pulled off a double bluff, and 5970 is going to easily match or better GTX580 with a significantly smaller and cheaper chip? With Antilles waiting to do more damage next month?

nope

 

[A.L.M.]

Looks like default is the lower TDP setting, or the newer drivers used here are a lot better.

Source?

 

[Erinyes]

Measuring up primitively with the only known factor, the PCI-Express x16 slot (89mm), I get the die to be 20,8 x 18,7 mm, which coincides with the rumored 389mm2.

Another reference could be the memory chips, I guess you could use either.

All we need now is a naked die shot of Gf100b/GF110 :smile:

Anyone know what the card length is? It looks pretty tight for most midi cases.

I think its the same as the 5870 which is 10.5". I've asked this question before and still haven't got a proper answer. Why is it that Nvidia with more complex PCB's, more memory chips, much bigger die package and greater power requirements(meaning more VRM circuitry) still manages to produce cards which match or are actually shorter than AMD's.(in the GTX 2xx series there was also an additional NVIO chip). AMD could help enthusiasts out by designing their boards better. The amount of unused space on the PCB is staggering

I haven't seen anything yet which said it's 2x full cayman. I guess though with the new dynamic clocking scheme AMD got a lot more room to play with clocks instead of disabling simds for power reasons, so it might make sense.

Whats surprising? HD 5970 also used 5870's at 5850 clocks. Dave explained the reason as well. There is leakage from the disabled parts as well so it makes more sense to enable the full chip but clock it lower.

No shit? That would be quite a coup after all the flak they got over tessellation.

I dont think you ranting about it constitues flak..

 

[mczak]

Whats surprising? HD 5970 also used 5870's at 5850 clocks. Dave explained the reason as well. There is leakage from the disabled parts as well so it makes more sense to enable the full chip but clock it lower.

I didn't say it's surprising (nor that I think it will be less simds), simply that noone has "confirmed" it will be all simds. It is quite true that generally enabling all units but clock it lower is better perf/power wise, however it need not necessarily be the case for all chips - and in fact I've got doubts about this for HD5970 since it scales quite badly with simd count (and much better with clock). It _could_ be beneficial to have a slightly higher clock but less simds (don't forget that this also boosts rop etc. performance, so for non-simd bound scenarios it would be faster).

 

[AnarchX]

Long awaited price war incoming?

Fudo: Radeon HD 6970 to cost 275 euro

With this HD 6800 cards should also see a reduction of 20-30€.

 

[UniversalTruth]

Long awaited price war incoming?

Fudo: Radeon HD 6970 to cost 275 euro

With this HD 6800 cards should also see a reduction of 20-30€.

That's amazing. I hope it's true.

 

[AnarchX]

That's amazing. I hope it's true.

Seems to be true:
http://www.coolpc.com.tw/phpBB2/viewtopic.php?p=276669#p276669

HD6970 ... 11998 元 ~299€
HD6950 ... 9198 元 ~230€

 

[Gipsel]

I didn't say it's surprising (nor that I think it will be less simds), simply that noone has "confirmed" it will be all simds. It is quite true that generally enabling all units but clock it lower is better perf/power wise, however it need not necessarily be the case for all chips - and in fact I've got doubts about this for HD5970 since it scales quite badly with simd count (and much better with clock). It _could_ be beneficial to have a slightly higher clock but less simds (don't forget that this also boosts rop etc. performance, so for non-simd bound scenarios it would be faster).

If you can't powergate the disabled SIMDs (which is extremely unlikely), it won't help too much. Clockgating doesn't do anything against leakage. If you have a sufficiently fine grained clock gating, it's virtually always better to have more SIMDs enabled as they won't cost more power if they don't do anything but they can potentially help the performance. It basically only make sense to do it if you want to decrease the difference between average and max power draw. But that can be done much better with the power containment feature of Cayman.

 

[Dave Baumann]

I think its the same as the 5870 which is 10.5". I've asked this question before and still haven't got a proper answer. Why is it that Nvidia with more complex PCB's, more memory chips, much bigger die package and greater power requirements(meaning more VRM circuitry) still manages to produce cards which match or are actually shorter than AMD's.(in the GTX 2xx series there was also an additional NVIO chip). AMD could help enthusiasts out by designing their boards better. The amount of unused space on the PCB is staggering

What is on the surface barely tells any story. None of NVIDIA's solutions are approaching memory speeds even close to our prior gen.

 

[Gipsel]

You mean like how AMD marketing "f*cked up" DP support -- another feature that a vanishingly small fraction of people care about?

Actually, "all" what's still missing are some decent math libraries. The hardware can do everything what is necessary in this respect. And the marketing has nothing to do with that

 

[Bouncing Zabaglione Bros.]

Seems to be true:
http://www.coolpc.com.tw/phpBB2/viewtopic.php?p=276669#p276669

HD6970 ... 11998 元 ~299€
HD6950 ... 9198 元 ~230€

AMD are going to sell a ton at that RRP. £225 for the 6950 and £280 for the 6970. A quick check on Ebuyer shows 570s at around £300 and 580 at £425. Nvidia will have to cut prices.

 

[Mize]

Long awaited price war incoming?

Big grin!
6970 CrossfireX for $600?
580 GTX for $350?

I'll take either or both!

 

[trinibwoy]

I dont think you ranting about it constitues flak..

Wow, certainly wasn't expecting a fanboy response to such an innocuous (and factual) statement. AMD hyped tessellation for years yet were embarrassed by Fermi in that respect. Now they have improved their geometry performance significantly enough to be taken seriously again. Those are the facts. Deal with it

What is on the surface barely tells any story. None of NVIDIA's solutions are approaching memory speeds even close to our prior gen.

Yeah that is a bit of a mystery and I'd love to know the details behind it. AMD's involvement in the development of the specification can't explain it all but neither can sheer engineering competency.

 

[Erinyes]

6870's are down to $239 again at Newegg after weeks of being > 250. So im predicting a price of 299/399 for Cayman Pro/XT respectively

What is on the surface barely tells any story. None of NVIDIA's solutions are approaching memory speeds even close to our prior gen.

I agree with you there Dave, and with the new gen the gap is even bigger. But how much space do memory traces take on a PCB? If you take all the other components i listed, theres a lot more besides just memory speeds. I think my 4850 is quite a nice size(its a Sapphire own design which is 1" shorther than reference) A 9.5" card would be the outer limit of card size i can fit comfortably in my case without obstructing my cabling and airflow around the hard drives(And i have an ok case, a CM 690)

Wow, certainly wasn't expecting a fanboy response to such an innocuous (and factual) statement. AMD hyped tessellation for years yet were embarrassed by Fermi in that respect. Now they have improved their geometry performance significantly enough to be taken seriously again. Those are the facts. Deal with it

Save it, its been debunked enough in the Tesselation thread.

 

[Dave Baumann]

But how much space do memory traces take on a PCB?

How fast do you want it to run?