Hacker News Reader

Comment by PaulKeeble | Hacker News Reader

Oddly for GPUs its the other way around and has been for a while. Nvidia and AMD seem to start with the mid and big dies first before filling out the small ones later. Intel seems to be coming from small to big however so they may reverse the trend at least for themselves. But it could still be driven by yield issues where the margins are less good for smaller dies and they sell in much larger volume so things need to be working well to hit the volume of the market and still be highly profitable.

View on HN

> Nvidia and AMD seem to start with the mid and big dies first before filling out the small ones later.

For GPUs, yield is less of an issue... the chips are manufactured with the expectation of a number of cores of the many thousand small (in terms of silicon area) ones being defective - overprovisioning, basically. That allows them to simply bin the sliced chips according to how many functional core units the individual chip has.

In contrast, even the Threadripper AMD CPUs have only 64 large cores which means the impact of defects is vastly bigger, and overprovisioning is not feasible.

Current AMD chiplets come with a maximum of 8 cores.

Right, that's a different strategy again. If you're making a monolithic 64 core die and one of the cores is defective, and the next one down in your product lineup is the 48 core, that's going to make the 64 core model harder to stock (and maybe not worth aiming for at all, if your yields are bad enough that this happens often).

Meanwhile if you're making 6 x 8 core chiplets and one of those cores is defective, well that chiplet can go into a 48 core or be a midrange consumer cpu or something, and you'll just pick one of your many many other 8 core chiplets to go with the rest for the 64 core.

Ala: The 14/16 inch MacBook Pro making all full-die function models built to order. I do find it interesting that they did a 7 core GPU version of the M1 much like the A12X/A12Z binning. I wonder if getting full function M1 was ever the intent.

Probably because the beefier GPUs for the server or compute market are way more profitable for Nvidia than high-end gaming GPUS, which are also way more profitable than chips Nvidia might make for laptops or low end gaming.

Reply on news.ycombinator.com