NVIDIA GF100 & Friends speculation

[KimB]

I don't think you're following me. Whatever scenario you can come with it will always be worse if the 512SP part is completely missing in action.

I think we understand what you're saying. The problem is you haven't argued why a launch of a non-existent part would actually help nVidia's bottom line.

 

[KimB]

What refresh? There's just 1 rumorsite which posted that there would be "fermi 2" during the fall, that's it

I don't think it takes a genius to recognize that there will likely be another iteration of the architecture coming quite soon.

 

[RobertR1]

In regards to the 512 part, they could certainly do a paper launch as Ati did with 5970 just for PR purposes. I'd imagine they will. Just make enough for reviewers that way you can boast about your flagship product and use the trickle down marketing benefits from it.

 

[aaronspink]

I don't think it takes a genius to recognize that there will likely be another iteration of the architecture coming quite soon.

What are they going to increase the die size?
What are they going to rent a time machine to go to 2011 when TSMC has they next gen process available?

Minor tweaks to existing designs on their existing process tend to be called respins.

 

[KimB]

What are they going to increase the die size?
What are they going to rent a time machine to go to 2011 when TSMC has they next gen process available?

Minor tweaks to existing designs on their existing process tend to be called respins.

The thing is, the delays that affected the soon-to-be-released part are unlikely to have effected the refresh parts. It is likely, therefore, that the refresh parts will come in rather short order (either this fall or next spring).

 

[3dilettante]

That would depend on whether Nvidia pulled resources from other projects to help push Fermi forward.

Any derivative product that was counting on Fermi to help iron out potential design issues and production kinks would be delayed since Fermi hasn't ramped.

 

[KimB]

That would depend on whether Nvidia pulled resources from other projects to help push Fermi forward.

Any derivative product that was counting on Fermi to help iron out potential design issues and production kinks would be delayed since Fermi hasn't ramped.

With past launch delays from nVidia, this hasn't been the case, and the refresh part has come in short order.

 

[Squilliam]

A3 went to the fab according to rumors in early December; it seems like they started mass production within February which makes sense. A hypothetical release in March makes sense for all the above and there doesn't seem to be any time for another metal spin (what for anyway?) let alone a full re spin which you'd rather expect somewhere in mid summer and not as early as March.

March for A3? That sounds alright. I wonder how much volume a late launch entails? Do we expect 30k, 60k or 120k boards at launch? They've been fabbing the chip for a while now but its hard to say how many wafers have actually gone through.

In addition, how ought the market respond? Ought people wait it out until a B variant arrives with faster clocks and full(er) availability of the 512SP variant? They played that game with the GT200 architecture, so why not expect the same for Fermi? If prices are expected to come down and performance come up why not wait?

 

[trinibwoy]

I think we understand what you're saying. The problem is you haven't argued why a launch of a non-existent part would actually help nVidia's bottom line.

I thought I made that very clear in my first post. The absence of a fully fledged 480 will reflect poorly on the 470 for several reasons. One, people will be hesitant to jump until they see what the full thing can do and two it's just bad press. The exact opposite of the halo effect will take place. The halo effect emanates more from reviews and opinions and not from whether it's in stock or not.

 

[XMAN26]

If this rumor is true, then I would tend to think that the most likely scenario would be that they would do one of a couple of things:
1. Go ahead and launch with a top-end part somewhat overpriced to compensate for the low volume. The reasonable-priced part will be a 448SP part.
2. Launch with a 448SP part as their flagship product, holding onto good 512SP parts until they have enough volume down the line for a more full release.

My reasoning is basically that at the current time, they're already late, and they're losing potential sales for every further month that it takes to release the GF100. So they're likely to release whatever they can release in volume as soon as they can.

Obviously I do hope that this rumor turns out to be wrong and the 512SP parts are available in volume (which might be the case if, for instance, nVidia has been saving the good 512SP parts for a little while).

They plan to go ahead and launch with the 448SP part. The reason for the respin/B1 is because 99% of what they have been getting off the line has been 448SP parts. As it stands right now, if they leave things as they are, the 512SP part will be a PE part only and most likely never see store shelves. This is NOT what they want, so they are going for a B1 sooner than anyone would expect.

 

[Sontin]

They plan to go ahead and launch with the 448SP part. The reason for the respin/B1 is because 99% of what they have been getting off the line has been 448SP parts.

How is this possible without a process problem?

 

[XMAN26]

How is this possible without a process problem?

Layout problem that didn't show up in testing but showed up during manufactoring. I dont know the process involved from design to tapout to making final silicon, but something somewhere along the way was messed up and didn't show up til final silicon was completed. Given it hasn't fixed itself in A3s, it would seem to be a layout issues causing the failures of 512SP parts.

 

[trinibwoy]

A layout problem that's solved by disabling two SM's? What kind of layout problem is that....

 

[XMAN26]

A layout problem that's solved by disabling two SM's? What kind of layout problem is that....

As isaid, i don't know the process involved from design to TO to silicon, but something isn't right somewhere if they are having huge issues getting 512SP parts but are getting 448SP ones in quanities.

 

[ShaidarHaran]

As isaid, i don't know the process involved from design to TO to silicon, but something isn't right somewhere if they are having huge issues getting 512SP parts but are getting 448SP ones in quanities.

Yeah, it's called "it's hard to get a 3+ billion transistor chip pushing the reticle limits of the process to yield significant quantities of chips without disabling some functional units".

 

[Jawed]

It's called making a rod for your own back

Jawed

 

[Sontin]

Yeah, it's called "it's hard to get a 3+ billion transistor chip pushing the reticle limits of the process to yield significant quantities of chips without disabling some functional units".

They did it with the G80 and GT200. I don't see a reason why they can't produce enough 512 chips for a real product.

 

[Kaotik]

I don't think it takes a genius to recognize that there will likely be another iteration of the architecture coming quite soon.

I'd expect "quite soon" to mean next process at TSMC, but didn't they cancel 32nm, putting 28nm as next bulk process, meaning 2011 at the earliest for those?

Of course they could do some sort of refresh like G80 > G92, but not the tradititional refresh type, like GT200 > GT200b or X1800 > X1900

 

[Rangers]

I'm kinda not sure what the big deal is about the 512SP part. It's only 12.5% more units/performance.

Just doesnt seem to be something all that critical. Consider it like the same ~10% performance improvement as a mid cycle refresh. Nothing like an all new part.

 

[XMAN26]

I'm kinda not sure what the big deal is about the 512SP part. It's only 12.5% more units/performance.

Just doesnt seem to be something all that critical. Consider it like the same ~10% performance improvement as a mid cycle refresh. Nothing like an all new part.

Because the rumor is, the 512SP part is supposed to have 200+ mhz higher core clocks(if that is true) over the 448SP part. Not to mention 48ROPs vs 40 and 384bit bus over 320.