Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
As said, people can misinterpret what Charlie is saying all they like, but thats not what he said.
THey can physically make Fermi cores, but if you're looking at 10 off a wafer, 9000 hot lot wafers(probably at an increased cost, but call it the normal $5k), then you're looking at an insane cost per core. That will likely increase with salvaged parts, but not enough. You're still looking at not far off a full 5850 in cost JUST to get the core out the oven, and thats just not good for business.
Nvidia can package up all 90k cores, at £5 for all it wants, it can call them the cheapest card ever that kills AMD. The problem is, there will still only be 90k of them and they'll take a humoungous loss on them. $5k x 9k = $45,000,000. No matter what or how they sell them, if they can't haul in $45million, then they've made a loss.
Charlie hasn't said they can't make a working Fermi, just that they've got a broken core, which, with a yield at sub 10%, is what the entire industry would call broken. Thats what he's talking about, and yes, it seems rather funny that Nvidia have been doing basic respins, if the first lot come back, over power, with a sub 5% yield, its time for a MASSIVE respin, IE the full relaying out of the core, not a basic spin that might add a few percent in yields and maybe a couple percent down in power usage.
The problem is, if they wanted to sell 2million parts, as AMD will likely sell of the 5850/5870 by the time the next gen is out, you're talking about, well, they literally can't put enough wafers through, or take that kind of loss.
As I've hinted at before, things will get WORSE for them on 28nm, yes TSMC's 28nm is highly likely to be much better than their 40nm, but compared to GF, it will be a disaster. Due to some simple designs for the process, take 3billion transistors on TSMC's process and the same amount of transistors on the GF process, its set to be about 20-30% smaller at GF< with less leakage, much less power, and higher clock speeds largely due to the previous two reasons.
The GF100 is broken, because even with a major respin, the 40nm process will still suck for a 1.5Ghz part it wants to be, with a die size so big, with so much leakage and being so power hungry. A full respin might offer 100% increase in their current yields, but that would still be up to 20% max, which is still awful.
The next gen version will still be too big, too hot, too power hungry, too late and not hit the clock speeds the architecture needs, because 28nm is going to have most of the problems 40nm will have, and being next gen its smaller but will want to use double the transistor count anyway, so the die size will stay the same problematic size.
Make it a car analogy, say you were trying to design, a I dunno, Ford Ka, small, compact and cheap, your plan is to sell at £10k, unfortunately you ended up with a Ferrari, great car that it is, at £150k, no matter how many you make, when you can only sell them for £20k, you're utterly screwed. It doesn't make a Ferrari bad, but as a product, its unmanufacturable and broken...... thats what Charlies saying, and it would take well, a deluded fool to disagree with him.
He's not saying there won't be cards this month, there might not be, but there will be some, at some point, it won't be many, it might not be that expensive. Because if they are limited to $45mil in wafers, and really won't bother with any more, make a loss of $10mil, or $45mil, not a huge difference to them. AT least if they sold them all at £200, they could keep promising new stock all the way till next gen and stop quite a lot of people going AMD.
It will be interesting to see what they do though, thats LOTS of options for them to spin, salvage, they might just be mental enough to keep producing lots at a huge loss, who knows right now.
