NVIDIA GF100 & Friends speculation

[Groo The Wanderer]

Yeah, that is kind of what I thought. Most insiders I talk to basically talk of hot clock and half hot-clock. I hadn't heard anyone internally talk of the base clock at all, and when they are talking yields, they just mention the hot clock. Top bin is way low for A3, but on the up side, it isn't yielding at all. Oh, never mind, that isn't an up side.

-Charlie

 

[3dilettante]

Just theorizing:

I wouldn't suppose the base clock would be pushing the limits of TSMC's process, while it appears the hot clock does.
In addition, it looks like the base clock region is in the center of the die, which might keep things uniform.
Possibly, those units are also composed of lower-leakage and slower gates whose thicker dimensions would offer more margin against variation.

A combination of the previous would relegate the global clock region to a secondary factor in parametric yield issues that the hot clocks would appear to be suffering from.

(edit: the half-hot clock region might have an unfortunate problem if the hot clock drops too low and forces it below the global clock)

 

[Ailuros]

I was told about a year ago that the clock targets were 1500 for the hot clock, and 750 (obviously) for the half hot clock. Does the base clock run anything anymore?

-Charlie

If those were truly the clock targets from the beginning then they might not be too far from those in the end. As 3dilettante said ROPs and L2 cache are supposed to run at core clock. A rough estimate puts the core clock somewhere on the 630-650MHz level; considering they now have 48 ROPs instead of 32 on GT200 it sounds like a healthy increase in maximum theoretical pixel fillrate. For rasterizing of course the pixel fillrate could be exactly on the same level, but that doesn't sound like an issue either.

I'm still personally worried about the bilerp rate of GF100.

 

[XMAN26]

If those were truly the clock targets from the beginning then they might not be too far from those in the end. As 3dilettante said ROPs and L2 cache are supposed to run at core clock. A rough estimate puts the core clock somewhere on the 630-650MHz level; considering they now have 48 ROPs instead of 32 on GT200 it sounds like a healthy increase in maximum theoretical pixel fillrate. For rasterizing of course the pixel fillrate could be exactly on the same level, but that doesn't sound like an issue either.

I'm still personally worried about the bilerp rate of GF100.

1500-1600 is still the hopeful target hotclock for the GTX480(512SP part). GTX470 will be in the 1250-1300 range with a core clock in teh 450-475 range unless they have change plan for the 448 part due to yields of teh 512 part.

 

[GZ007]

1500-1600 is still the hopeful target hotclock for the GTX480(512SP part). GTX470 will be in the 1250-1300 range with a core clock in teh 450-475 range unless they have change plan for the 448 part due to yields of teh 512 part.

Why should they have much higher clocks on the 512 SP than the 448 SP when they have the same binning problems just with several times less working parts. That wouldnt be to economical.

 

[Alexko]

Why should they have much higher clocks on the 512 SP than the 448 SP when they have the same binning problems just with several times less working parts. That wouldnt be to economical.

Well, if the target for the GTX 480 is 512SPs and ~1500MHz, then it makes sense for Nvidia to salvage parts that don't have 512 functional SPs as well as parts that can't hit ~1500MHz; hence the 448-SP, ~1300MHz GTX 470.

 

[GZ007]

Well, if the target for the GTX 480 is 512SPs and ~1500MHz, then it makes sense for Nvidia to salvage parts that don't have 512 functional SPs as well as parts that can't hit ~1500MHz; hence the 448-SP, ~1300MHz GTX 470.

Maybe but than they will end with 10 470gtx cards for a single 480gtx card.

 

[Alexko]

Well that's better than 5 GTX 470s for a single GTX 480 (and the same number of 480s), isn't it?

 

[kyniskos]

That presumes they have 2 equal production issues though, both clock and functional units. Really does not sound promising.

 

[Ailuros]

1500-1600 is still the hopeful target hotclock for the GTX480(512SP part). GTX470 will be in the 1250-1300 range with a core clock in teh 450-475 range unless they have change plan for the 448 part due to yields of teh 512 part.

I thought the difference between core and 1/2 hot clock is in the 12-14% range? If yes and I take the hypothetical 1250 of a 470 I get a 1/2 hot clock of 625 and an estimated core clock +/- 550MHz.

Your minimum 450MHz core clock for the 470 gives 18 GPixels/s while a modest 640MHz core clock for a 480 gives 30.7 GPixels/s. Are you sure the hypothetical roughly 40% difference there isn't a tad too large especially if the 4th rasterizer hasn't been disabled?

 

[Sontin]

Well, if the target for the GTX 480 is 512SPs and ~1500MHz, then it makes sense for Nvidia to salvage parts that don't have 512 functional SPs as well as parts that can't hit ~1500MHz; hence the 448-SP, ~1300MHz GTX 470.

I thought that they don't have real 512 Parts...

 

[XMAN26]

I thought the difference between core and 1/2 hot clock is in the 12-14% range? If yes and I take the hypothetical 1250 of a 470 I get a 1/2 hot clock of 625 and an estimated core clock +/- 550MHz.

Your minimum 450MHz core clock for the 470 gives 18 GPixels/s while a modest 640MHz core clock for a 480 gives 30.7 GPixels/s. Are you sure the hypothetical roughly 40% difference there isn't a tad too large especially if the 4th rasterizer hasn't been disabled?

The 450-475 core for the ROPs is propably due to trying to salvage as many parts as possible which would point to clock problems across the whole die. But, to be honest, even I have been very skeptical of those core clock numbers from my source as they dont make sence as you noted. Given everything else tho, it could be true. We should know whats what in about 4 weeks or so.

Heres a thought that maybe they should have gone with. GTX 460-480 with 460s having the 450-475 core clocks, 470s having 500-550 and the 480s being 600-650.

 

[XMAN26]

I thought that they don't have real 512 Parts...

No, they have SOME 512sp parts. whether or not we'll see any on launch day as PE pieces for reviews is the million dollar question tho. I wouldn't expect to see any tangable 512SP parts on store shelves til mid-late summer.

 

[Sontin]

No, they have SOME 512sp parts. whether or not we'll see any on launch day as PE pieces for reviews is the million dollar question tho. I wouldn't expect to see any tangable 512SP parts on store shelves til mid-late summer.

And AMD had a few rv740 chips with specification in march 2009 - but we never saw anyone of them in a notebook.
The GTX480 is not a real product with 0,5 chips from every wafer. And that means that all the talk about clocks and configurations of the cards are bogus. They will sell Tesla with 48 ROPs and 384bit interface. And on the desktop they stay with a non existent GTX480 and a low clocked GTX470 with 320bit?

 

[XMAN26]

And AMD had a few rv740 chips with specification in march 2009 - but we never saw anyone of them in a notebook.
The GTX480 is not a real product with 0,5 chips from every wafer. And that means that all the talk about clocks and configurations of the cards are bogus. They will sell Tesla with 48 ROPs and 384bit interface. And on the desktop they stay with a non existent GTX480 and a low clocked GTX470 with 320bit?

Look man, I'm simply sharing info I've been given. I can't help you dont like the info. But if B1 fixes whatever the issue is, the 512 parts should start to become a reality by mid-late summer. Lets just hope all these delays hasn't killed off the mid and lowend parts til 2011. That would simply be to long for NV to not have DX11 parts across the board against ATI.

 

[digitalwanderer]

No, they have SOME 512sp parts. whether or not we'll see any on launch day as PE pieces for reviews is the million dollar question tho. I wouldn't expect to see any tangable 512SP parts on store shelves til mid-late summer.

Not in any way sarcastic or intended to cause offense, but what does "PE pieces" mean? I couldn't figure it out.

 

[DavidGraham]

Not in any way sarcastic or intended to cause offense, but what does "PE pieces" mean? I couldn't figure it out.

Press Edition !

 

[Groo The Wanderer]

Well, if the target for the GTX 480 is 512SPs and ~1500MHz, then it makes sense for Nvidia to salvage parts that don't have 512 functional SPs as well as parts that can't hit ~1500MHz; hence the 448-SP, ~1300MHz GTX 470.

From what I have seen, all companies make the cut down part based on both fused off shaders and down clocks. This lets them make a single bin for the 'leftovers' rather than two or three downbins, something that would be a marketing pain in the ass. ATI, Nvidia, Intel and AMD all do it this way when applicable.

-Charlie

 

[FrameBuffer]

The low full part (assuming 512 Core) rate is still quite astonishing at this point in the game, I don't think anyone would doubt the abundance of lower (470) parts at launch with very limited availability of the 480, which at this rate looks like another 6800 GTX 512MB Ultra Extreme "Press Edition" meant to capture the lead and create a spot as NV's halo part. Given the flack over such past parts (5900 - 6800/7800 GTX and 800XT PE/5970) I'm not sure how forgiving the masses will be this time around. I have no doubt that loyalist to each camp will parrot the company mantras but I certainly hope that the press starts to (for once) do their jobs and actually address the issue of availability rather than stick to company issued talking points in hopes that they'll continue to keep their "favored" status and receive products and services that come with such status.

 

[dkanter]

The low full part (assuming 512 Core) rate is still quite astonishing at this point in the game, I don't think anyone would doubt the abundance of lower (470) parts at launch with very limited availability of the 480, which at this rate looks like another 6800 GTX 512MB Ultra Extreme "Press Edition" meant to capture the lead and create a spot as NV's halo part. Given the flack over such past parts (5900 - 6800/7800 GTX and 800XT PE/5970) I'm not sure how forgiving the masses will be this time around. I have no doubt that loyalist to each camp will parrot the company mantras but I certainly hope that the press starts to (for once) do their jobs and actually address the issue of availability rather than stick to company issued talking points in hopes that they'll continue to keep their "favored" status and receive products and services that come with such status.

You're being awfully generous to the 'press' with your expectations. A few folks might consider availability, but frankly most don't.

The gap between knowledge in industry and press is quite amazing.

David