[gary_0]

Depending on who you ask, binning is segmentation. Generally demand isn't going to exactly match how the yields work out, so companies often take a bunch of perfectly good high-end chips, nerf them, and throw them in the cheapo version. You used to be able to (and still can, in some cases) take a low-end device and, if you'd won the chip lottery, "undo" the binning and have a perfectly functional high-end version. For some chips, almost all the nerfed ones had no defects. But manufacturers like nVidia hated it when customers pulled that trick, so they started making sure it was impossible.

[creato]

> You used to be able to (and still can, in some cases) take a low-end device and, if you'd won the chip lottery, "undo" the binning and have a perfectly functional high-end version.

For the purposes you tested it, sure. Maybe some esoteric feature you don't use is broken. NVIDIA still can't sell it as the higher end SKU. The tests a chip maker runs to bin their chips are not the same tests you might run.

I'm sure chip makers make small adjustments to supply via binning to satisfy market demand, but if the "technical binning" is too far out of line from the "market binning", that's a lot of money left on the table that will get corrected sooner or later.

edit: And that correction might be in the form of removing redundancies from the chip design, rather than increasing the supply/lowering the price of higher end SKUs. The whole point here is, that's two sides of the same coin.

[wtallis]

Disabling cores that have 100% passed QA is quite commonplace, especially for chips that have been on the market for over a year and thus are being built with yields as mature as they're going to get.

Artificially restricting supply of high-end chips and increasing supply of mid-range chips by disabling fully functional cores is how chip makers preserve their pricing structure. Without doing this, market pressures would force prices down on high-end chips and cause lower bins to mostly disappear from the market, leaving the product line with lower overall margins and a PR nightmare every time a new generation launches with pricing reset back to the initially high levels.

As a rule of thumb: if a chip product line goes a whole year without having new SKUs show up with a higher percentage of cores enabled or higher clock speeds for the same core count, then the manufacturer is artificially restricting supply to make more of the lower-bin parts than naturally occur in the fab output.

> And that correction might be in the form of removing redundancies from the chip design, rather than increasing the supply/lowering the price of higher end SKUs.

Those two courses of action take place on completely different timescales. Disabling cores and other binning tricks can be implemented in no more than a few months. Adding a new chip with a different number of copies of the same IP blocks takes well over a year. Removing redundancy within an IP block (eg. by having fewer spare SRAM blocks for a cache of a fixed capacity) isn't going to happen within a single chip generation.

In the semiconductor world, corrections of any kind tend toward "later" rather than "sooner".

[MarkusWandel]

Dumb question (maybe): I'm aware that "tester time" is very expensive for advanced integrated circuits. Could it be that disabled cores are actually "unknown" i.e. probably good, but money was saved by not even testing them?