Discussion
Moore's Law is dead.
Intel introduced the 64 bit "Core 2" processor in 2005 so laptops could access more than 4 gigabytes of RAM.
Moore's Law's 18 month doubling time says the following 21 years should have 14 doublings to 64 terabytes of RAM in each $2000 laptop.
That's why silicon valley execs went crazy: they no longer have free growth. Their business model doesn't WORK. They've done regulatory capture, brute force spending inflating a snake oil bubble, and nothing's working.
@landley
Moore's law is indeed largely gone. But it was about transistor density, not about the capacity or price (Moore never claimed things would get cheaper!) of the resulting systems, and it never really applied to ram.
@jannem The original 1965 Moore's Law paper was about RAM. Intel was founded to produce DRAM.
https://www.landley.net/history/mirror/interviews/Moore.html
RAM doubled like clockwork (largely due to industry collusion in production schedules) for decades, I wrote about it in 2003:
http://www.catb.org/~esr/halloween/halloween9.html#id2867629
And again in 2006, that time with example products/ranges sourced from an extensive collection of old computer magazines:
http://catb.org/~esr/writings/world-domination/world-domination-201.html#id248066
@landley
I stand corrected - he was talking about ic components in heneral! And I agree Moore's law no longer fully holds (and will eventually slow to a crawl).
But my greater point is still valid: it talks about manufacturing density. Not about the capacity, speed, or price of the resulting product. The end of Dennard scaling already decoupled that. Smaller components aren't able to run faster, or made cheaper.
@jannem Chip designs hit a clock speed wall circa y2k, then went wide (pentium M was basically just a pentium III with the racy logic cleaned up and megabytes of L2 cache on die), then they went SMP, then the put frame buffers and eventually GPUs on die to soak up the transistor budget (AMD bought ATI in 2006).
The mainframe->minicomputer->microcomputer->smartphone migrations muddied the waters a lot (I've given talks about those), but when I say moore's law is dead past tense, I mean it.
@landley
Yep, that was the end of Dennard scaling. I think people underestimate the impact of that.
Your RAM observation got me thinking:
It's easy to forget that Moore's law was an economic observation. It's says it's worth keeping that development pace, because the revenues will make it worth the cost.
Both sides of that equation have to hold. Each new process node costs more. So the benefits side has to increase in proportion. And that's stopped happening for two reasons:
🧵
@landley
One is the speed wall. That was mitigated in part by multicore systems and specialized silicon. But ultimately you're limited by total power - dark silicon throttling are there for a reason. And if more transistors doesn't make your daster faster, why pay more?
The second - and I think, as important - is non-chip limits. When a smartphone or laptop is fast enough, why pay more? The mass market has an economic limit.
🧵
@landley
And that's where I realized perhaps memory is extra interesting.
Our limits for memory isn't capacity. You could easily (well, not today) build a machine with 128G. Our compute nodes all have 500G each. But for most use-cases you *don't need it*.
And memory is mostly latency limited. Faster RAM chips don't matter when there's a 200ns latency to the CPU.
The demand side for memory hit Moore's law earlier and harder than for other components. And that's why it's clearest there.
@jannem I use memory as an example because it's easy to explain, and because I have those researched charts to pull out when challenged. :)
Screen resolution made it to "4k". (640x480, 1024x768... Doubling was 18 month-ish for a long time but in pixel count, and with specific step increases. I mentioned the collusion...)
Data storage continued to expand for 5-10 years longer than RAM (8TB usb drives were briefly affordable retail). Gigabit ethernet (1999) was "good enough" for endpoints.
@jannem The thing is, there _was_ demand for more memory (if nothing else why didn't chromebooks all come with 32 gigs ram), but the price, speed, power consumption, and heat dissipation didn't improve fast enough to make it worth it. So marketing talked it _down_ instead of up.
(All this is complicated by the sideways migration: "Blade servers" were laptop components used in servers. Raspberry PI was phone tech in a tiny desktop style case/motherboard...)
@jannem Processors hit significant limits years before memory did.
Keep in mind an S-curve bending down doesn't mean things don't continue to increase. The slowdown went something vaguely like 1.5 year douling, then 2 years, then 3 years, then 5 years. And suddenly in 12 years when you should have had 8 doublings, you've had 4 and are 16x behind and nobody noticed because marketing was frantically boiling frogs in front of you. And the NEXT 5 years you go _backwards_ and blame "the pandemic"
@landley @jannem What happened next with marketing was a pivot to snake oil: first cryptocurrency miners, then LLM platforms, all of which soak up transistors like crazy. And a huge sub-population of the pyramid growth obsessives peeled off to behavioural advertising and thence to stuff that is de facto malware (only legal because if you're rich enough you can buy the laws you need to work with).
I'm typing this on a laptop that came out in 2012. It was high-end then and I've upgraded it as far as it can go, but no piece of hardware that's NINE DOUBLINGS behind should still be usable, let alone slightly more powerful than a modern chromebook.
That's why silicon valley's gone septic. Everything they did was pantomime while photolithigrapic die size shrinks did all the heavy lifting. Now the S-curve's bent back down again there's no more growth to be had, and they're in SCREAMING DENIAL.
@landley 2012, that's almost perfect. I wrote this in 2014: https://astoundingteam.com/2014/11/02/monetization-value-subtraction/
Install bonfire.cafe
Get the full app experience
