That certainly makes sense. I didn't know if there's any liklihood that now that Nvidia knows what the 5870's performance looks like, if the clock speeds they envisioned for the A2 silicon Fermi ended up not being enough in testing to beat it, or at least beat it by enough. And with the 40nm process problems, it was decided to go for another respin.
The fact that they haven't shipped out the build kits confirms earlier reports about Nvidia wanting A3 silicon maybe. Though I'm not sure if the dates add up, or if this delay is due to something else.
http://vr-zone.com/articles/-rumour...to-be-a3-revision-q2-2010-/8014.html?doc=8014
Steve I don't know whether you find much value in
these kinds of
posts, but perchance you do then I figured I'd extend the analysis to include Fermi and maybe some of the possible business decision rational would fall thru as an unavoidable consequence of the maths.
Making some rough estimations, if alternative numbers are preferred let me know and we can iterate thru the "if...then" consequences as well, but we know Cypress is
334mm^2 and a consistent (IMO) estimate of Fermi has it weighing in around 530mm^2.
We also know both chips are fabbed in the same fab, so both chips experience the same D0 (matters to yield).
Building off the Cypress vs. Juniper example in the AMD 58xx supply thread we arrive at the following functional yield entitlement correlation comparison (represents upper-limit w/o invoking harvesting/fusing):
Calculating the max number of dies per wafer using:
We arrive at 167 dies for Cypress and 99 dies for Fermi (not factoring in yield losses).
Factoring in functional yield losses we are expecting net die per wafer to decline to
~90 for Cypress and
~38 for Fermi.
These numbers could be higher depending on the actual D0 of the fab and the methods of harvesting and fusing (converting those 5870's into 5850's, etc) and can also be lower if D0 is larger and as we account for parametric yield (binning, etc).
What does this mean for cost per die? 40nm wafers go for around $7k, all the obvious and previously discussed caveats go into this number so I won't rehash them here, so we'd be expecting Cypress to cost ~$78 and Fermi to cost ~$184 based on the myriad of assumptions we made in this analysis.
Now we know both AMD and NV would like to see their chips fetch them 50% GM, that means
Cypress selling for $160 and
Fermi selling for $370.
(that's $200 more for the Fermi GPU than the Cypress GPU right there, just getting the chips to the AIB's)
The AIBs are going to pay that, plus more for the memory, power components, PCB, assembly/packaging/shipping and they want their gross margins too. Then on to resellers, Newegg, ZZF, who want their GM as well.
So yeah, pretty quickly we can see where TSMC's 40nm yields can pretty much make or break the market opportunity for a Fermi-size chip if it needs to be priced within $200 of a Cypress-based SKU.
That D0 needs to be around 0.10 (or less) such that the price differential between Cypress and Fermi declines below $100 at the AIB point in the chain of supply.
Facebook
Twitter