Hacker News Reader

Comment by spankyspangler | Hacker News Reader

DRAM scaling has come to an end.

Memory prices will only move sharply down again if a different memory technology is developed that can be manufactured with finer features.

View on HN

Would this also mean fabs are effectively much cheaper since they don't become obsolete every couple of years?

Not really. Fabs don't become obsolete for 20-30 years because there is ample demand for processes that are decades old. Profit margins drop but by the time they drop significantly, the fab investment is generally paid off and it becomes a gravy train. AFAIK Intel's last plant to close was D2 in Santa Clara - it was built in 1989 and closed in 2009 after 30 years. Their earlier Chandler, AZ plant closed after 20 years.

Sure, oversimplification in my question I guess - the returns on older fabs don't go to zero, but their products fetch a margin that's a function of how close to bleeding edge the process is. But the question still stands.

(btw fab site lifecycles don't usually correspond 1:1 to fab equipment generations, they tend to get refitted, this was also the case in Intel's D2 which had an old process by the time of retirement but not 30 years old)

The big costs are the variable costs per chip though, which is why technology shrinks have delivered such great cost savings for the end product despite the fact that every shrink has required ever greater outlays in development costs and capex for the manufacturing plant. The saving in the variable costs are so large that they pay for all that and the price drops.

So even if you have entirely paid off an old fab or bought one outright for ~$0 at a fire sale, you probably won't be able to make DRAM at a hugely lower overall cost. Slightly lower, perhaps.

The figures I found put capital costs as the highest slice of the pie and growing faster than materials or labour costs [1] [2]

[1] http://smithsonianchips.si.edu/ice/cd/CEICM/SECTION2.pdf p. 5)

[2] https://www.icknowledge.com/news/Technology%20and%20Cost%20T... p. 9

I could well be wrong, sorry I should not have tried to make such a strong statement I don't actually have much knowledge of recent technology nodes.

In your first link I'm struggling to see it demonstrate what you say. Actually that's more what I'm familiar with.

Obviously it's quite old though. The second link does seem to show the trend going to equipment costs and trending upward indeed. Although it doesn't seem very detailed, I suppose that first graph on the left on page 9 looks like mostly fixed costs in orange. Thanks for the link.

> DRAM scaling has come to an end

Can you elaborate? First time I hear this.

CMOS process scaling is running out of steam which is generally going to make performance/capacity/etc. improvements in things dependent on semiconductors increasingly difficult.

Others have explained better than I could. Here's some a search turned up

https://semiengineering.com/dram-scaling-challenges-grow/

https://blocksandfiles.com/2020/04/13/dram-is-stuck-in-a-10n...

Reply on news.ycombinator.com