Hacker News new | past | comments | ask | show | jobs | submit login

We're almost at (or are already at) the stage where the packaging of the chip costs more than the computer inside it.



we're two orders of magnitude past that stage. a ryzen 7 is ten billion cmos transistors for a hundred dollars, ten nanodollars per transistor. so how much does a minimal computer cost at ten nanodollars per transistor?

the intersil im6100 was a pdp-8 cpu in 4000 cmos transistors, and the 6502 was comparable, but that's with no memory other than the cpu registers. for a useful microcontroller you probably need about 8192 bits of instruction memory and a few bytes of ram, so let's round up to 16384 transistors for a whole computer. an 8051, with built-in instruction eprom and 128 bytes of ram, was 50k. the arm2 without ram was 27k. an avr, with built-in ram and flash, is 140k. https://en.wikipedia.org/wiki/Transistor_count

at that price, counts of 16384 to 131072 transistors work out to 0.016¢ to 0.13¢. but the cheapest computer you can buy today is a padauk pms150c https://www.lcsc.com/product-detail/Microcontroller-Units-MC... which is 4.2¢ in onesies (and 2.43¢ in quantity) with 64 bytes of ram and 1024 13-bit words of one-time-programmable prom for the program https://free-pdk.github.io/chips/PMS150C. that's 150× more; in the day of moore's law doublings every two years that would have been 14 years, but now it's probably longer. (incidentally this same blog looked at them a few years ago https://cpldcpu.wordpress.com/2019/08/12/the-terrible-3-cent...)

ergo we've been at the stage where the packaging of the chip costs more than the computer inside it since about 02009

(obviously the ryzen 7 cpu costs a great deal more than its packaging, though, because that's what you have to do when you're competing on computrons per dollar rather than gpios per dollar. in theory for 2.43¢ you should be able to get 2.4 million transistors, enough for about 300 kilobytes of rom or a 50 kilobytes of sram, or half that together with a 486 or a quad-core arm3. presumably padauk is not doing this because they're using long-obsolete semiconductor process nodes, which is also why their chips are so power-hungry)


What a complicated way of saying 'yes' ;)


i thought it was an interesting enough question to merit deeper analysis; perhaps someone else who thinks so will read it


I agree with OP/OP/OP. Thanks for the initial point, breakdown and the entertainment :p

Can't wait for us to start seeing authentic 3d silicon tho. Gimme that 3mm^3 cube of magic. And beyond that, computronium.


I have had Microchip FEs say as much to me.

Will be interesting when chiplet technology gets to the point where multiple chips are installed in a plastic package (there are thermal expansion/contraction issues that make this a hard problem but still ...)


If and when the cooling problem is solved and we can just stack layers at will computronium will truly have arrived. Adding another dimension will unlock hardware potential that we can only dream of today.


chiplets still lose area to scribing and dicing




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: