R520: 32 PS pipelines? Any truth to this?

IMO Nvidia has a better product right now compared to r420... It has almost the same performance as r420, sm3 and the "famous" accelerated video encoding / decoding.... My question is: why launch a superior product before ati?

 

I just read the whole thread and I think Dave is saying that the rumors last generation were much closer to the truth than the rumors this time around. That leads me to believe that R520 will not be 32 pipes for sure and the transistor count is much too high at 350 million. I also sense a very disturbing lack of enthusiasm for R520 from many other posters that I think know something.

If ATI is again put in a position to be stuck with a generation that requires MAXIMUM (100 MHz more than Nvidia) clock speeds to compete, then there are very serious problems at ATI. Gross margins have been stalled at 34% for ATI the past several quarters while Nvidia has seen dramatic gross margin expansion. I am completely convinced that the early shortages of R420 (especially AGP) were the result of poor yields on the 130 nm low-k process. This shortage lasted many months and only when ATI moved some SKU's to 110 nm did the supply become less tight. While capacity may have played an issue as well, yields were without question an issue as I can find not other explanation of ATI being stuck at the low end of their 34-38% GM guidance for so long.

IF and I do mean IF ATI is in a similar position as R420, on a new process, facing yield issues and the prospect of having to push clock speeds to theoretical maximums to compete against Nvidia then there is something very wrong at ATI. Was R3xx a one trick pony from Orton? If R520 doesn't take back the initiative from Nvidia, recapture market share, be available in mass quantities for PCIe/AGP and improve gross margins then it is a failure. As such, Orton will have shown himself to be a great engineer of R3xx and then to be a shitty CEO that was unable to duplicate his earlier success.

What happened to ATI's great idea of introducing new generations with a midrange product on the newest cutting edge process and then riding the older process for the flagship? I would mush rather see R520 on 110 nm and maybe the x700 replacement on 90 nm. To have to work the bugs out on the flagship SKUs with a new process smacks of what went so horribly wrong with R420.

 

The generally considered best card (and correct IMO when it generally plays at the same settings or better) is the X850XT PE. If Nvidia want that mind share they need a new product. SLI may have offset this to a point, but Crossfire has managed to set the new 3dmark05 record (dont think its on the ORB yet). Having the G70 on the market next month will allow them to keep the top of the ORB

Im surprised you would mention Nvidias video chip as so far it seems to have been of minimal benefit, and its not even completly functional on the 6800s (or has that changed with their PCI-E ones?)

 

Actually ATi's best card is the X800XL and has gained them tons more mindshare than the PE ever did. I thought the video problems went away with the debut of the PCIe 6600 line but I've never actually seen a report of it working.

 

Market timeing issues come into play as well. The builders have cycles too. You've made the investment, time to make some revenue. Maybe NV would have preferred --all else being equal-- to see R520 first, but that doesn't mean they are willing to wait a couple extra months to make it happen. Imagine how well (NOT) that strategy would have worked out for ATI with R300 if they'd decided they needed to see the near-mythical juggernaut that NV30 was supposed to be before releasing their own part. . .

That's part of what people tend to forget about NV30 --not just its own problems, but how severe the disappointment compared to what was expected.

 

Indeed, thats definitly got them more market share than the X850 series, probably mind share as well considering SLI clouds the X850 XT PEs top position.

TBH even if Nvidias video thingy worked as expected I still dont see whats so great about it, encoding is not something most users do that often and decoding is plenty fast on current cpus taking into account the content they are likely to come across.

 

But the process choosen for r520 is one of TSMC's high volume ones is it not?

So when it ramps up, and yields get decent, the darned things might acctually be readily available at MSRP?

 

For what I remember hearing, they have triple number of 90nm low-k nodes compared to 130nm low-k nodes

 

I predict ~300M transistors and 32 pipelines total - not 32 pixel pipes, but 24 pixel plus 8 vertex

 

Completely wrong; I predict 16 + 16 + 16 +16 + 8 = 72 pipes. Since nowadays one can call virtually anything in those quad based designs a "pipeline".....

 

Ati has had the XBox 360 GPU to worry about lately and considering the vastly different architeture (unified shaders, EDRAM...), they probably had a lot of the workforce working on that part. So i wouldn't worry so much about Ati even if the R520 ends up with the same position as the R420. Could even be a bit worse before i start to worry.

 

The hardware guys should be just about done, shouldn't they? Tho I think the driver guys might be still a few months from giving their full attention to pc stuff. Hopefully a good bit of this pays off next year on the pc side as well. Orton has referred to 2005 as "an investment year" that pays off in 2006.

 

Keeping the lead?

 

I'm thinking that Ati will be in good shape when the R500/XBox 360 architecture arrives in PC land. Though i guess that at this point, we don't really know if that architecture (unified shaders) are the way to go but it certainly sounds more interesting then the RSX.

 

a.) As far as I can tell the Marlborough guys have been pretty much working in their own silo since even before they started work Xenos so I don't think that will have affected them that much in terms of the DX9 development path - possibly it may be Revolution that would have more of an impact on current tech, depending on how close or different it is.

b.) I think you can bet you bottom dollar that the basic shader/threading architecture will be the core of R600.

 

Maybe I mised something, it's been a while since I've been able to keep up my reading assignments here (busy: work, skiing, school, etc), please forgive me if this has been beaten to death and discounted already.

I'm just wondering; what ever happened to the idea of 24 'standard' pixel pipes and then 8 'adaptable/flexible' (or as they're called in the R500 'dynamically scheduled') pipes

It seems that would still be possible with a relatively low transistor count compared to the normal 32 pixel pipelines with the usual vertex compliment.

This would also build on their experiences from designing the R500 without jumping the gun too much on putting the hardware before the software. It might also be a wise use of the room afforded that the switch in processes offers. Of course it's still a risk of wasted transistors if the implementation is limited. It'd also explain some of the confusion over transistor count and architecture.

With the mediocre return of 110nm non-low-k it would seem natural that the move to 90nm low-k would be attractive for a top end part; while being perceived as risky as the NV30's transition to 130nm, it seems to be more 'necessary' in this case because of the available options from TSMC (or did they add 110nm low-k and 90nm non-low-k capacity while I was away too? ).

Anywhoo, really just curious what ever happened to the 24+8/8 idea that seemed so logical at one time?

 

I dont believe that was ever my point. If I make it sound like that I apologize. The point I am trying to make is when it comes to intensive games that will really make use of DX9 both cards will be rendered useless.

In Unreal Engine 3 does it matter if the R300 renders at 4fps vs NV30s 2?

Sure, it is twice as fast but completely worthless. Thus the DX9 performance of both cards IMO ended up being completely irrelevant in the end.

Basically the entire life of the cards while in their prime was a bunch of synthetic benchmarks being pitted against each other.

 

It's a waste of time/money/resources to do this to 8 fragment pipelines and leave the other 24 alone. All or nothing.

 

I think it will more likely end up being say 20 vs 40fps in 640x480, playable vs unplayable. UE3 isnt going to be THAT demanding or even current SLI setups would be in the teens I heard rumours that the 9800 pro would be the minimum specification.

Anyway the r3xx have already proved their dx9 ability in half life 2, tomb raider and to a less extent farcry where if you want to run at 1024x768 no fsaa you should be able to reach very high quality

 

So the news of R520's death (performance-wise, that is) have been greatly exaggerated, haven' t they?

At least that's what the INQ says and we all know the INQ is always right.