[VC] AMD Radeon HD 9000 Series Launching in October?

[Lanek]

Huumm, dont forget AMD have offtly use risk production.... and they was really good at it ( mean the "so called" risk production" was a sucess ( on engeener side ) " ... they have offtly use it . it will not be the first time ..

Risk mean hazrdous in term of yields, this dont mean hazardous in term of cost.... The GPU team is really good to push this, and dont forget, if the 20nm is ready tomorrow, they work on it since 3 years allaready .. ( this dont mean i say i think they use 20nm )

[xray]

The "stable through 2013" slide only included Q3, not Q4.
As for the Power Color reps...who says they weren't lying or just misinformed? Seems unlikely to me that they would give away such an important piece of information just like that.

Whistling past the graveyard?

Of course if it was a rumor that NV had pulled it off....

[Zloyd]

Hmm im quite confident they will beat 780 ...... dont think anyone cares about SuckerTan. 7970 has 680 comfortably ATI engineers push and repeat the performance !!! (8950 insane bang for buck card incoming hope i can get 2))

[DefStar]

For anyone in the know, when can we expect full cover water blocks to be available for the 9970 and 9950??? Thank you.

[Final8ty]

Lol ^

[DefStar]

Lol ^

That is, when does E.K. or Swiftech usually release a block? And what are you laughing at...dude? There is no doubt in my mind that AMD will release a card faster then Nvidias and will still cost less even after buying the water block. Also, I just got my first Loop this weekend...so i'm a little restless.

[Final8ty]

I'm only laughing because the cards are not out yet and still away's off.

You've got the itch real bad

[LordEC911]

That is, when does E.K. or Swiftech usually release a block? And what are you laughing at...dude? There is no doubt in my mind that AMD will release a card faster then Nvidias and will still cost less even after buying the water block. Also, I just got my first Loop this weekend...so i'm a little restless.

They usually get the card specs early enough that they can have blocks out right around launch, usually between a month or two for full coverage.

[Lanek]

Ek should get a full waterblock when the card release or in the weeks who follow.. ( i dont remember have buy an AMD card at release without having my EK full cover block, the next week who have follow ).. they got the design schema really fast and they adapt their full cover block really quick with it...

[DefStar]

I thought there was suppose to be a frame pacing driver released today or did I miss something?

[MaddMuppet]

It's just being tweaked a little. Have patience my young padawan

[BenchZowner]

Whenever AMD launches a product, hardware or software doesn't matter, without any delays the last 2 years, a major earthquake happens somewhere...

[Andrew LB]

whats ive heard, will be launched sooner then october and highest one-gpu model beats Titan

That's what is commonly referred to "wishful thinking". The only way they'd get close to Titan's performance is if the HD 9xxx series was manufactured using the 20nm process. And as we all know, TSMC is still 6 months away from even beginning "low-volume production". Which means high-volume production will be another couple months after that.

Transcript of the shareholder meeting where this info came from can be found here: http://www.semiwiki.com/forum/conten...4nm-track.html

[PedantOne]

no, Hawaii is still on 28nm, but more SPs and higher clocks will reach Titans performance level

[LordEC911]

That's what is commonly referred to "wishful thinking". The only way they'd get close to Titan's performance is if the HD 9xxx series was manufactured using the 20nm process. And as we all know, TSMC is still 6 months away from even beginning "low-volume production". Which means high-volume production will be another couple months after that.

Transcript of the shareholder meeting where this info came from can be found here: http://www.semiwiki.com/forum/conten...4nm-track.html

As I've said before, 20nm is NOT the only way they can match/beat Titan.

[Tao~]

As time passes 20nm seems more likely.

[zalbard]

As I've said before, 20nm is NOT the only way they can match/beat Titan.

I find it unlikely they will go for a chip as big as GK110. So I agree with Andrew LB.

And I don't believe in magical huge increases on largely the same architecture...

[LightSpeed]

I find it unlikely they will go for a chip as big as GK110. So I agree with Andrew LB.

And I don't believe in magical huge increases on largely the same architecture...

It might not beat Titan, but it should be close. Titan is around 35% faster than 7970GE and higher clocks + cores + tweaks should be able to give it ~30% more performance. Should be similar to a GTX 780 while consuming similar if not less power.

Some say that the power consumption would be way too high in the above scenario but I dont think so. Aside from the core/process tweaks, they have a big room for improvement with regard to turbo power states/voltages and from what ive seen from Bonnaire, its pretty significant and can only get better.

[MaddMuppet]

As time passes 20nm seems more likely.

Highly unlikely, unless they wait until 1st or 2nd quarter 2014.

[Budwise]

Whatever it is I just hope its not $1000.

[tajoh111]

It might not beat Titan, but it should be close. Titan is around 35% faster than 7970GE and higher clocks + cores + tweaks should be able to give it ~30% more performance. Should be similar to a GTX 780 while consuming similar if not less power.

Some say that the power consumption would be way too high in the above scenario but I dont think so. Aside from the core/process tweaks, they have a big room for improvement with regard to turbo power states/voltages and from what ive seen from Bonnaire, its pretty significant and can only get better.

Tweaks at this points are just that, tweaks. They hardly lead to big jumps because its not a new architecture. Adding more shaders is typically less efficient at this point because theirs a bottle neck in the architecture somewhere. The small difference between a 7970 and 7950 at equal clocks show this. Higher clocks and more cores = higher power consumption. Increasing utilization of existing shaders is going to increase power consumption further like what happened to the 6970 vs 5870.

30% is no small task at the same node, with no new architecture and GCN being kind of a energy pig as it is. Adding 25 percent more shaders an increasing clocks another 200mhz is going increase power significantly. I think adding any more shaders will require AMD to add more ROPS which will increase the size of the chip significantly.

Its no coincidence that Nvidia clocked their gk110 line lower. Making a bigger chip causes significantly more power consumption, more so than performance often because of leakage. It just happens that Nvidia's boost technology helps salvage a big more performance while it also seems Nvidia is less reliant on clocks this gen.

AMD brought 40 percent more performance initially over the 7970(not counting driver boosts). They had a new architecture and a new node. Although more came with drivers. GCN 2 is likely to be similar to gcn 1 and thus driver boosts won't be that much.

Getting the next 9970 to perform 30 percent faster while generally conditions are against it(gcn being big as it is and consuming as much as it is, no new node and no new architechture) make it something difficult to pull off, unless they want to make a 450mm2 plus chip.

If the 7970 was around 300mm and didn't consume so much power when overclocked, I could see AMD doing this. But the 7970 is 365mm and drinks power like a glutten at 1200mhz.

[LordEC911]

~440-480mm2 and roughly the same clocks, maybe a bit lower, should be enough.

Improved front end, more/stronger ROPs, 30-40% more shaders.

I'm not sure why you think ROPs take up a bunch of room, that was more true when ROPs were linked to the memory controller which isn't the case with GCN.

Also, with a mature process they can potentially increase transistor density, remove some redundancy that was needed for Tahiti and used the knowledge they got from the rest of 7x00 series and the 8x00.

Will it be easy? Nope.
Is it impossible? Nope.

[LightSpeed]

Tweaks at this points are just that, tweaks. They hardly lead to big jumps because its not a new architecture. Adding more shaders is typically less efficient at this point because theirs a bottle neck in the architecture somewhere. The small difference between a 7970 and 7950 at equal clocks show this. Higher clocks and more cores = higher power consumption. Increasing utilization of existing shaders is going to increase power consumption further like what happened to the 6970 vs 5870.

30% is no small task at the same node, with no new architecture and GCN being kind of a energy pig as it is. Adding 25 percent more shaders an increasing clocks another 200mhz is going increase power significantly. I think adding any more shaders will require AMD to add more ROPS which will increase the size of the chip significantly.

Its no coincidence that Nvidia clocked their gk110 line lower. Making a bigger chip causes significantly more power consumption, more so than performance often because of leakage. It just happens that Nvidia's boost technology helps salvage a big more performance while it also seems Nvidia is less reliant on clocks this gen.

AMD brought 40 percent more performance initially over the 7970(not counting driver boosts). They had a new architecture and a new node. Although more came with drivers. GCN 2 is likely to be similar to gcn 1 and thus driver boosts won't be that much.

Getting the next 9970 to perform 30 percent faster while generally conditions are against it(gcn being big as it is and consuming as much as it is, no new node and no new architechture) make it something difficult to pull off, unless they want to make a 450mm2 plus chip.

If the 7970 was around 300mm and didn't consume so much power when overclocked, I could see AMD doing this. But the 7970 is 365mm and drinks power like a glutten at 1200mhz.

I mostly agree with what you said, but this was their first shot at GCN and a completely revamped architecture. There are bottlenecks that can be ironed out and as I said, Bonnaire showed what could be done simply with a better implementation of voltage/power states. Also, Tahiti consumes a lot of power while overclocked/volted but a new architecture coupled with process tweaks will surely enable it to operate at lower voltages.

I wonder if AMD are willing to make a ~450mm^2 architecture though.

[PedantOne]

ok, it can be maximaly 2560 SPs, 40 ROPs, same number of TEX units. Same memory capacity, bus, clocks maybe
with the same GPU clock as HD 7970 GHz, it can be +20-25 percents performance above Radeon HD 7970 GHz in best case

[LordEC911]

ok, it can be maximaly 2560 SPs, 40 ROPs, same number of TEX units. Same memory capacity, bus, clocks maybe
with the same GPU clock as HD 7970 GHz, it can be +20-25 percents performance above Radeon HD 7970 GHz in best case

I'm thinking (the biggest they would do):
More shaders- 44CUs or double the SIMDs and TFUs per CU and go with ~24CUs(3072ALUs and 192TFUs) [depends on how much a diff the stronger FE makes]
Better frontend- the 4ACEs and 3geometry rumor thats been floating around sounds like a decent increase.
Slightly more ROPs- I don't see them going with 48ROPs unless they move to 512bit w/ 6ghz GDDR5(unlikely) or starve/make the ROPs more bandwidth efficient.
If they use 7Gbps GDDR5, that would offer ~27% more bandwidth allowing them to do the 40ROPs, a 25% increase.

Core clocks would be under 1ghz, probably around 900-950mhz(one would hope).

That should fit in <500mm2(imo 440-480mm2) and I would assume stay around 250-265w TDP.

That should roughly match Titan.