NVIDIA Fermi: Architecture discussion

[Silus]

Jokes aside, it's a good reminder for all of us that the true miracles of nature are lightyears apart from silly hw or any other product out there. I'm still proud that we "created" rather a Cypress than a Fermi, since she arrived 13 days earlier in perfect health

I don't think anyone (actually, everyone not a fanatic) ever thought that the miracles came from pieces of silicon

 

[Silus]

Contrary to Fermi, I have 2 2900 XT/1G sitting in my archive.

You better. That thing was released over 2 years ago

 

[Ailuros]

I don't think anyone (actually, everyone not a fanatic) ever thought that the miracles came from pieces of silicon

You don't have to be a fanatic follower of whatever brand to get too absorbed by certain things outside of true miracles. It was just a reminder how awefully silly it is to get emotional over a silly piece of silicon and yes there are times where I'm not completely innocent either.

 

[Jawed]

Fermi is being considered "late" from the launch of the HD 5800s.

Er, no, late from the launch of W7.

40nm at TSMC carries quite a chunk of blame. But NVidia has basically been late with everything since G84/86, so I wouldn't be too hard on TSMC.

Arguably some Evergreen GPUs are late, too. Cedar/Redwood should be out there now. I can't help thinking that Cypress was meant to be a May/June launch.

Jawed

 

[Silus]

You don't have to be a fanatic follower of whatever brand to get too absorbed by certain things outside of true miracles. It was just a reminder how awefully silly it is to get emotional over a silly piece of silicon and yes there are times where I'm not completely innocent either.

Enough of that anyway! Enjoy your baby and keep her away from geeks that talk about pieces of silicon...you are the exception to that rule obviously ^_^

 

[Ailuros]

Er, no, late from the launch of W7.

40nm at TSMC carries quite a chunk of blame. But NVidia has basically been late with everything since G84/86, so I wouldn't be too hard on TSMC.

Arguably some Evergreen GPUs are late, too. Cedar/Redwood should be out there now. I can't help thinking that Cypress was meant to be a May/June launch.

Jawed

I'll say it again: IMHO NV needs to reconsider its strategy and release a performance part first with high end and lower end parts to follow. Let's just accept for now that it still wasn't possible with this generation, but I don't see there being an excuse for future generations especially since chip complexity is getting more and more critical.

 

[Silus]

Er, no, late from the launch of W7.

Okay, then that makes Fermi even less "late".

40nm at TSMC carries quite a chunk of blame. But NVidia has basically been late with everything since G84/86, so I wouldn't be too hard on TSMC.

Yeah, TSMC, at a time where they should be showing their customers how they should keep using their fabs (given the upcoming competition), are fumbling with the 40nm process.

Was G92 late ? I don't recall it being late, but then again NVIDIA had no competition at the time, so there wasn't much to track...

Arguably some Evergreen GPUs are late, too. Cedar/Redwood should be out there now. I can't help thinking that Cypress was meant to be a May/June launch.

Jawed

Who knows, but ATI is certainly suffering from TSMC woes. They are just better than NVIDIA, since there are at least some new cards out.

 

[Silus]

I'll say it again: IMHO NV needs to reconsider its strategy and release a performance part first with high end and lower end parts to follow. Let's just accept for now that it still wasn't possible with this generation, but I don't see there being an excuse for future generations especially since chip complexity is getting more and more critical.

IMO that's more or less what I think they'll do. They will most likely release the GeForce 380 (full chip, though with some HPC specific features disabled/removed) along with a new chip with half or a bit more than half the units of the 380, to cover the mid-range segment (GeForce 350 maybe ?). They may re-capture the performance crown and thus "mindshare", but the ultra high-end is not the most profitable segment, so they need to have a performance part aswell.

 

[mapel110]

40nm has been a problem for nearly all chips. I don't think just lowering the transistor count would help that much in such a case.

 

[rpg.314]

40nm has been a problem for nearly all chips. I don't think just lowering the transistor count would help that much in such a case.

Chip complexity and consequently design effort scales superlinearly with tranny count, everything else being equal. Process troubles are due to TSMC in this case, and no fabless ihv can fix that, apart from making sure that they didn't break any design rules.

 

[neliz]

Realism finally takes control of Fuad

In January, Nvidia should have the final samples and a limited number of Fermi Geforce GT300 chips, but the launch might take place later.

If pushed, Nvidia might launch Fermi Geforce in late January as the final chips should be there by then, but real volume shipments should start towards end of Q1 2010.

The most realistic availability date is March 2010, and again only if everything goes right. Judging by our previous information, Nvidia delayed its plans by more than one, if not two quarters.

TweakTown's take

Every day I try to somehow snag a Fermi sample from one of the many companies we work with here at TweakTown. The problem is for the first time over the weekend we heard the 'M' word come out of a partners mouth in regards to the release. What's the 'M' word? March!

January seemed a pretty set date from most partners when it came to the release of the new NVIDIA based graphics card, the problem is though that for the first time we've now heard a February / March timeline. You can't help be feel worried that the time is slipping further and further away.

 

[Silus]

Doesn't surprise if the other "article" is true too:

40nm ATI / Nvidia shortage to last whole Q1 10

http://www.fudzilla.com/content/view/16686/1/

So it's bad for ATI, worse for NVIDIA...but it's especially bad for us.

I doubt that they don't release the first Fermi based products in Januray though. There will be extremely limited quantities, but they will exist in the consumer market at least.

 

[fellix]

I wonder, what is the prospect of NV being forced (by the circumstances) to release a cut-down GeForce Fermi SKU with 256-bit interface and scaled down number of multi-processors (12 for instance) -- new IC or salvage part?
Obviously the 3B transistor count monster is a too large bite for TSMC.

 

[Arty]

I wonder, what is the prospect of NV being forced (by the circumstances) to release a cut-down GeForce Fermi SKU with 256-bit interface and scaled down number of multi-processors (12 for instance) -- new IC or salvage part?
Obviously the 3B transistor count monster is a too large bite for TSMC.

GT212 part duex.

 

[trinibwoy]

What possible reason could there be for ATi's parts to still see low yields all the way into Q1?

 

[Silus]

I wonder, what is the prospect of NV being forced (by the circumstances) to release a cut-down GeForce Fermi SKU with 256-bit interface and scaled down number of multi-processors (12 for instance) -- new IC or salvage part?
Obviously the 3B transistor count monster is a too large bite for TSMC.

That's what I speculated a few pages back (possibly the GeForce 360), but I wouldn't put that as a cut down version of the full chip. The cut down version should be a GeForce 370 with 448 or 480 Stream Processors on a 320 bit memory interface or something along those lines.

 

[trinibwoy]

I'll say it again: IMHO NV needs to reconsider its strategy and release a performance part first with high end and lower end parts to follow. Let's just accept for now that it still wasn't possible with this generation, but I don't see there being an excuse for future generations especially since chip complexity is getting more and more critical.

Well that sorta assumes its trivial to make a small chip and then make a bigger chip based on the same design. Not even the CPU guys do that. So even if they were to get the smaller chip out faster it might take them that much longer to do the bigger variant. Of course, this is based on an assumption that scaling up a design is harder than scaling it down.

 

[gkar1]

I'll say it again: IMHO NV needs to reconsider its strategy and release a performance part first with high end and lower end parts to follow. Let's just accept for now that it still wasn't possible with this generation, but I don't see there being an excuse for future generations especially since chip complexity is getting more and more critical.

It must be quite satisfying for AMD see their design philosophy pay off. Smaller chips that take a lot less time to develop and cost less to manufacture.
At the same time it must be frustrating having that pay off delayed by circumstances out of their control. They designed and taped out their chip months ahead of the competition but now the manufacturer is holding them back. I never bought the rumour and nonsense talk of AMD not ordering enough wafers from TSMC.

Realism finally takes control of Fuad

You know you're in trouble when your puppet mouth piece loses faith. There, there dear Fraudo. There, there.

I wonder, what is the prospect of NV being forced (by the circumstances) to release a cut-down GeForce Fermi SKU with 256-bit interface and scaled down number of multi-processors (12 for instance) -- new IC or salvage part?
Obviously the 3B transistor count monster is a too large bite for TSMC.

That's assuming they even have any chips (regular or salvage parts) to begin with. All signs point to them having neither at this point. Even if they did, time to market would still be 3-4 months away since partners would have to be notified of these changes and they'd have to start printing new boxes, ramp up new PCB designs, etc, etc. Unless they are desperate enough to reuse the more expensive reference PCB design with its extra layers just to launch something. If I were them I'd force Nvidia to subsidize the cost of the more expensive PCB if that's what Nvidia shovels for them to sell.

 

[LordEC911]

What possible reason could there be for ATi's parts to still see low yields all the way into Q1?

No reason, just bad info or averaging.
AMD/ATi's only problem with TSMC is capacity.
Yields are not as good as expected for Cypress at this time but have been improving steadily and Juniper has better than expected yields.
http://www.driverheaven.net/news.php?newsid=322
http://www.theinquirer.net/inquirer/news/1563212/amd-hd58xx-shortages

 

[Groo The Wanderer]

What possible reason could there be for ATi's parts to still see low yields all the way into Q1?

That the yields suddenly go down from where they are now perhaps? Fudo is horribly wrong about the yields, they are MUCH better for ATI than he is saying. I have some strong suspicions of what NV's are, internally described as "horiffic", and a much better idea of what ATI's are.

As I said here, ATI did their homework, and now have a good handle on the process, and how to work with it. NV was arrogant an tried to intimidate physics into doing what they wanted by yelling a lot. It didn't work.

I would bet that Fudo is parroting back NV FUD to try and deflect sales from ATI when the pipes start filling in a couple of weeks. Everything I hear says that article is complete bull.

Also, you should frame that argument in the context of NV blaming TSMC for all their woes lately. That saves them from having to take blame themselves, something that is against company rules, just like honesty.

-Charlie