Mostly when talking about yields they will be refering to how many parts on that wafer are fully working, every last transistor working.
Binning brings the yield up after the fact, the yield is how many top end cores come off the line.
4% is fairly believable to be honest, based on how many cores they got for A2 silicon(rumoured to be under 10 per wafer, with that Fermi core we saw being marked as number 7iirc). Considering the Tesla product has been changed at the high end from 512sp's to 448sp's, the yields on the highest speed parts are low enough they simply don't believe they can make them in suffient quantity to sell in the professional sector.
But it could quite easily mean what you might refer to as, for instance people say the 58XX series cards, the X denoting any old number, instead of an X they used a "-" , so essentially 4X% , anything from 40-49% yields.
But really this is the problem, far to many reporters here yield and think of a single thing.
TSMC didn't have any drop in wafer yields at all, though it was widely reported they did, they had a drop in the plant output, from 9k to 6k wafers, but same yield on the wafers themselves.
IT could mean anything basically, and real yields are never publically released so we'll never know past a good guess.
Considering A2 was rumoured SO low, I very much doubt a theoretical 40-50% yield would mean for the top end part, if they were only getting 40-50% of the wafer as working cores at all, even heavily binned ones, those are some terrible numbers.
D.P almost exclusively bigger cores always have worse yields, simply because even if only one core failed on every wafer, which is basically a given on any process, a bigger core means less cores per wafer, so one not working gives you a bigger percentage yield drop.
Its perfectly possible that a bigger chip could get lucky and have say 10 faults causing it to fail all on one chip with the rest fine, and a wafer of smaller chips could also have 10 faults but one each on ten cores and so a worse yield. In reality, big chips always have a worse yield.
AMD have rumoured yields of 60% or so, though I would guess that might be an average, with 57XX cores bumping the average up significantly, wouldn't be surprised to see the 58XX versions with a lower yield than that.
Anyway the only really important info is production in 3rd week of feb. Thats, well, not good news, if production starts in 3rd week of Feb, you won't even start to get finished wafers back till 2nd week of April, after which you're looking at another two weeks going from Nvidia, to the board makers, to the distro's then the shops. That will be without ANY time to build up a small supply. Normally they'll start production and wait 2-4weeks to build up a small stockpile. If they release from the first second possible, basically 8 weeks after production starts, you're looking at basically very few cards going out on a weekly basis. If yield are ultra low than that supply is even worse.
The strange thing is, they were due A3 silicon back basically 1st week of Jan, thats another 6 weeks till production is due to start. Thats a heck of a lot of testing they seem to be wanting to do and a long time waiting to put it into production.
Whats also obviously possible is in production from whoever told them, means the getting the cores back from the Fab part and "producing" the finished cards ready for shipping is what they are counting as production, which is an odd way to put it. Which would roughly tie in with final finished A3 silicon is back, tested and they've approved production.
Which could mean limited supplys mid March, but weeks/months before theres decent supply as they certainly don't have time to build up a stock of chips before release.
