AMD: R9xx Speculation

[CarstenS]

http://www.gpu-tech.org/content.php...sellation-punch-does-the-Geforce-GTS-450-have

I think for real world workloads it would be more than sufficient if AMD removed the 3 clocks per tessellated triangle bottleneck instead of adopting Nvidias approach with decoupled triangle setups. That seems to be quite difficult to do - both from an engineering standpoint (there AMD shouldn't have a big problem) but also with regard to die size and maybe power - which would be more crucial if AMD will pursue its current route.

What I don't understand from my own benchmarks though, is why the GF100s tessellators seem to be less efficient than the ones in the smaller chips. I don't believe that I already have run into the systems limit.

 

[Jawed]

I've got no idea what TessMark is doing, so I'll refrain from pondering the figures.

But in general a four-way split is going to suffer from Amdahl's law more than a two-way split.

Honestly I think tessellation is very poorly understood now. It's way more complex than GS, and that's barely understood.

 

[fehu]

I see rumors about the use of differential signaling gddr5+ in the 6XXX top models

 

[Silent_Buddha]

I thought that was pretty much what most of our guesstimations based on the leaked benchmarks were all about?

Given Napoleon's new performance hints, I still think Barts XT should be named HD 6850, btw: Too fast (and probably too big and too power hungry) to fall into the midrange HD 67xx category, yet not fast enough to actually beat the current HD 5870 (and thus hard to present as the new "HD 6870") ...

Your speculation perhaps. Most of the rest of us are still convinced Cayman is 68xx.

It makes absolutely zero sense in any scenario for 6850 to be approximately the speed of 5850 and for 6870 to be approximately the speed of 5870.

Just look at all the uproar when 5870 and GTX 480 were on average ~50% faster than the cards they replaced. And now you're somehow convinced that cards that replace their previous incarnation should remain roughly the same speed as the card they replace?

Makes no sense.

Now when you get to cards below the performance/enthusiast class chips (IE - x7xx and below), well that's always been all over the board depending on how well the IHV's have been able to scale down the top end chips. Sometimes it's a large increase in performance and sometimes there is very little to no increase in performance. But that's not a reflection on how the top of your lineup should be performing.

It's possible perhaps that a Barts chip may make it to 6830, but I remain highly skeptical of that. 2 cards per speed bin should be plenty fine to fill all price segments if performance scales well going down the line. In other words rather than last gen where there was a large gap between 5850 -> 5770, it'd be better if there was a smaller gap between those both in price and performance.

That eliminates the need for an x830 card for anything other than salvage reasons in the case of non-optimal yields. Anything more than 2 cards per speed segment just serves to clutter your lineup overly much, IMO. Hell, personally, I think 2 cards per speed segment is already too cluttered.

Regards,
SB

 

[Mianca]

@Silent_Buddha

What I meant by my first phrase was that Cayman XT performing roughly at the same level as HD 5970 is what most rumors and leaked benchies hinted at so far.


As for the second part of that post I made - sorry if I evoked the impression that I thought my own view on the naming subject was the "general view" of this forum. I know it isn't

I should have made it more clear that this second thought was actually just a more "personal" side-note, not an extension of the "our guestimations" phrase.

Thanks for your views on this, anyway. I'll admit that I can be stubborn as a mule - so don't waste your time on me :smile:

 

[neliz]

No, it is the new HD 6850 and HD 6870. But, yep, performance wise grin it may be the new HD 5970.

y.. no?

 

[mao5]

http://www.chiphell.com/thread-125814-1-1.html



Barts Pro>GTX460
Barts XT>HD5850

some one in Dell and Asus always leak the wrong info to chiphell

 

[trinibwoy]

I think for real world workloads it would be more than sufficient if AMD removed the 3 clocks per tessellated triangle bottleneck instead of adopting Nvidias approach with decoupled triangle setups.

Yep, Fermi's tessellation is way overengineered / wasteful for current applications. AMD can achieve adequate performance with a more straightforward approach.

What I don't understand from my own benchmarks though, is why the GF100s tessellators seem to be less efficient than the ones in the smaller chips. I don't believe that I already have run into the systems limit.

Well there a host of other things happening in the frame that would be affected by the different balance of resources in GF100/GF104/GF106. Sort of impossible to draw any conclusion based on a single resolution. What you would need to do is see how each chip's tessellation performance scales with resolution/AA etc. Also, your numbers are calculating per SM performance, not per tessellator and GF100's SM's are weaker than on the derivative parts (meaningful for hull and domain shading).

Edit: On my 460 @ 715Mhz performance at the insane setting ranges from 32fps @ 1440x900 0xAA to 27fps @ 1920x1200 8xAA so it's not completely geometry bound.

 

[racca]

I thought that was pretty much what most of our guesstimations based on the leaked benchmarks were all about?

Given Napoleon's new performance hints, I still think Barts XT should be named HD 6850, btw: Too fast (and probably too big and too power hungry) to fall into the midrange HD 67xx category, yet not fast enough to actually beat the current HD 5870 (and thus hard to present as the new "HD 6870") ...

So you are saying 5770 *should've* been named "5850"? It can even beat 4870 plus DX11 capabilities.

What exactly does BartsXT have that makes it so special? It's nowhere near 5870. It has no new features, nothing that can be compared with DX10.1 to DX11 at least. I shall consider even 6750 could be a proper name in that regard. But 6770 should be more appropriate given it's an XT part.

 

[racca]

Evergreen was several weeks late, implying that AMD had a preference for the start of September.

Also, when did these next GPUs tape out? March?

EDIT: I'm wondering if Cayman has slipped by 2 months, putting it one month after Barts, reversing the order of last year's launch.

Weren't there talks about re-taping Cayman a few months back? That should push Cayman back at least 8-12 weeks.

As for the launch date, Juniper finished earlier than Cypress, so is Barts/Cayman. I don't see the reverse on the slipping part.

OTOH, if the dead line was set October (the same as last time). Then it makes perfect sense. The highest performance part (excl. dual chip) always comes first as long as it can make the deadline for marketing reasons. So Juniper was in no hurry as Cypress made it in time too. Cayman isn't so lucky, October hard launch is not promising. I think AMD would launch both Barts and Cayman in October with only Barts actually shipped to retail. Cayman would be a paper launch with possibly a shipment date slated and a few ES cards sent to be benchmarked.

 

[neliz]

So you are saying 5770 *should've* been named "5850"? It can even beat 4870 plus DX11 capabilities.

What exactly does BartsXT have that makes it so special? It's nowhere near 5870. It has no new features, nothing that can be compared with DX10.1 to DX11 at least. I shall consider even 6750 could be a proper name in that regard. But 6770 should be more appropriate given it's an XT part.

Thanks for filling us in here. could you give us the numbers of the benchmarks too? I kind of like to know them.

 

[fbuffer]

I think for real world workloads it would be more than sufficient if AMD removed the 3 clocks per tessellated triangle bottleneck instead of adopting Nvidias approach with decoupled triangle setups. That seems to be quite difficult to do - both from an engineering standpoint (there AMD shouldn't have a big problem) but also with regard to die size and maybe power - which would be more crucial if AMD will pursue its current route.

What I don't understand from my own benchmarks though, is why the GF100s tessellators seem to be less efficient than the ones in the smaller chips. I don't believe that I already have run into the systems limit.

Ok stupid question time... looking at the results of Tessmark and comparing the 5770 vs 5870 they seem to both drop linearly until extreme where both the 5870 and 5770 both score about the same (and insane being equal). This got me thinking.. IF the 5770 is almost exactly half of a 5870 (minus DP-FP) and two 5770s in crossfire get approx. 40% faster/higher scoring than the 5870 even given the clock rate differences. If two 5770s are double the tesselation setup vs a 5870 (?) then (heres a long awaited question) .. then would not two 5770s in Crossfire (effectively a 5870/50) represent an 5800 with improved/fixed triangle setup ?

I guess it would make more sense to clock the 5870 and 5770s equally then compare.. A guesstimation on my part, it would seem that two 5770s (or a "improved" 5800) would get double the fps then the "hobbled" 5800. I don't know.. just thinking out loud.

I hope that sort of makes sense... been u all night and my brain isn't exactly firing on all cylinders.

 

[Jawed]

Weren't there talks about re-taping Cayman a few months back? That should push Cayman back at least 8-12 weeks.

Honestly can't remember.

As for the launch date, Juniper finished earlier than Cypress, so is Barts/Cayman. I don't see the reverse on the slipping part.

We don't know how much longer Cypress spent from initial tape-out to retail in comparison with Juniper. Also the later launch of Juniper includes an element of stock-building, a period that is theoretically longer than that required for Cypress. But, again, no real numbers.

Does a sold-out halo card have a shinier halo than an in-stock one?

OTOH, if the dead line was set October (the same as last time). Then it makes perfect sense. The highest performance part (excl. dual chip) always comes first as long as it can make the deadline for marketing reasons. So Juniper was in no hurry as Cypress made it in time too. Cayman isn't so lucky, October hard launch is not promising.

I suppose there's mileage in the idea that Barts will produce higher margin than Cayman as this new family ramps up. Then there's the laptops question, where Cayman is essentially a no-go. The longer lead times for laptops bias things towards Barts or lower. Though there's apparently only 2 windows each year for new GPUs to go into new laptops.

Actually that brings up an interesting point. Is Barts too big/hot for laptops, as a replacement for Juniper?

I think AMD would launch both Barts and Cayman in October with only Barts actually shipped to retail. Cayman would be a paper launch with possibly a shipment date slated and a few ES cards sent to be benchmarked.

I suppose there's also the competition for 40nm production capacity brought about by Ontario/Zacate. AMD's already shown that halo GPU production will be sacrificed in favour of OEMs and their preference for lower end parts.

 

[Dave Baumann]

Where does this myth that Juniper finished before Cypress come from?

 

[SimBy]

Where does this myth that Juniper finished before Cypress come from?

Good question, maybe some have misinterpreted what Anand said:

"As an interesting aside, when AMD started sampling Evergreen cards to game development houses and other 3rd parties, they were Juniper based, and not Cypress based. The Juniper team was rather proud of this, particularly since Juniper came back from TSMC second. They also had less time to get their GPU up and working than the Cypress team did, since they had to wait on Cypress before being able to finish work on some elements. This is what makes AMD’s 6 month rollout all the more impressive, since it means the non-Cypress teams had less time to get their work done than they have in previous product cycles."

http://www.anandtech.com/Show/Index/2856?cPage=5&all=False&sort=0&page=1

 

[neliz]

Where does this myth that Juniper finished before Cypress come from?

http://www.anandtech.com/show/2856

Juniper was what they got before Cypress. hence stems "Juniper was finished before Cypress"

As an interesting aside, when AMD started sampling Evergreen cards to game development houses and other 3rd parties, they were Juniper based, and not Cypress based. The Juniper team was rather proud of this, particularly since Juniper came back from TSMC second. They also had less time to get their GPU up and working than the Cypress team did, since they had to wait on Cypress before being able to finish work on some elements. This is what makes AMD’s 6 month rollout all the more impressive, since it means the non-Cypress teams had less time to get their work done than they have in previous product cycles.

 

[Squilliam]

Where does this myth that Juniper finished before Cypress come from?

Possibly because people saw/heard of earlier samples with Juniper. For instance the early DX11 cards developers tested/created their games with were Juniper were they not?

 

[Dave Baumann]

They also had less time to get their GPU up and working than the Cypress team did, since they had to wait on Cypress before being able to finish work on some elements.

Read that carefully, it still say Cypress is in front...

 

[gongo]

I thought that was pretty much what most of our guesstimations based on the leaked benchmarks were all about?

Given Napoleon's new performance hints, I still think Barts XT should be named HD 6850, btw: Too fast (and probably too big and too power hungry) to fall into the midrange HD 67xx category, yet not fast enough to actually beat the current HD 5870 (and thus hard to present as the new "HD 6870") ...

Thanks but no thanks! By virtue of being in x7xx series...it will be priced in the value/perf range...don't give Dave AMD team any idea of renaming it to x8xx range....the first DX11 gpu batch were disappointing....high price...low than expected perf increase...still holding out on ma 4870 1GB.

 

[Silent_Buddha]

http://www.anandtech.com/show/2856

Juniper was what they got before Cypress. hence stems "Juniper was finished before Cypress"

The important part from that quote bolded...

As an interesting aside, when AMD started sampling Evergreen cards to game development houses and other 3rd parties, they were Juniper based, and not Cypress based. The Juniper team was rather proud of this, particularly since Juniper came back from TSMC second. They also had less time to get their GPU up and working than the Cypress team did, since they had to wait on Cypress before being able to finish work on some elements. This is what makes AMD’s 6 month rollout all the more impressive, since it means the non-Cypress teams had less time to get their work done than they have in previous product cycles.

As well the rest just follows from that. I guess people just focused on the first part (which cards were sent to dev houses) way too much and ignored the rest (which chips came first).

Regards,
SB