What we think about it:
YogurtclosetFeeling6 - "This title makes me feel dumb. I don’t know if this means it’s really good or terribly bad."
Sirisian - "https://semiconductor.samsung.com/us/support/tools-resources/dictionary/semiconductor-glossary-yield/"
alidan - "at a .2mm2 defect rate which is relitively high but was normal for a long long time and a high bar at others, amd 7900 xt/xtx gets gets over 150 gpus of working parts (this is math I did a while back so exact numbers are lost, I think it was 180 but may have been 140) and I believe a near 90% yield rate (if they use dummy dies on the 7900xt they get technically 100% because even dead ones are used as spacers) while nvidia for their stull gets about 30-40% of the wafers useable for that part."
jaypeeo - "Sounds good though I’m not up to date. QC has to be extensive with such a high precision and sensitive process so there are abnormally high reject rates vs more forgiving products. A chair can be high 90s. A surgery, well… depends."
VexingRaven - "Not sure you get to have a reject rate with a surgery lol. You get what you get and you make it work."
Elbynerual - "Or you DIE. Which, I guess, would be the reject rate."
lenski7 - "Or even better you're crippled, which is also not a fun option!"
loercase - "There are plenty of outcomes between success and death."
lightwhite - "Doctors can't save everyone, you know. Doesn't take medical malpractice to have a rejection rate."
VexingRaven - "Yeah but you don't go "nah this surgery wasn't good enough" and throw it out lol"
afcagroo - "It's actually meaningless. Describing yield without mentioning chip size says almost nothing. 75% is really, really good for a new process making some of the biggest chips. It's lousy if the chips are small or if the technology is very mature."
unskilledplay - "It's a meaningful number. No buyer cross shops between 130nm, 28nm and 4nm. There's no reason to normalize metrics to make these comparisons."
jjayzx - "There are old nodes used for plenty of things. Not everything needs the latest and greatest technology."
unskilledplay - "Right. Tons of new chips use old nodes. My point is that no customer would design a chip and then make a choice between 130nm and 4nm. Designs aren’t even portable between those nodes. So why would a metric that allows for a comparison be meaningful?"
jjayzx - "Ah yes, sorry. Got the wrong intent on what you meant before."
Hawk13424 - "We use a lot of older processes for automotive and industrial. Also analog. We use 130, 90, 16, and 5, all for some new designs."
afcagroo - "That's not my point. Obviously, no one is looking at such disparate technology nodes (in most cases). But people do want to look at things like how well a new technology or new fab is performing. And if you are shopping between foundries, you certainly want to make such evaluations. That's what that article was attempting to do - poorly. It was lazy writing targeted at people not in the industry, and implied that a 75% yield on it's own is great stuff. It is, for a full-reticle die with many metal layers. For a small die, no."
badabababaim - "Also doesn’t mention how much it costs per wafer. 4nm ain’t shit if it’s a million per wafer"
MINIMAN10001 - "We know that following the trend graphs cost per transistor is on the rise and the number of transistors per product is on the rise. So it'll cost a lot."
