> And yet we do it every ~18 months with semiconductors.

So, this is about to stop. Very soon. Quantum tunneling, yield rates, etc are all starting to be cost prohibitive.

<10nm stuff was already delayed. There's plans for 3nm. It's unclear yet whether 2nm will work.

Even if hypothetically it does, there's a very real limit - an atom is only 10x smaller than 2nm.

Sure. But there’s been questions about the viability of the next node for at least twenty years. If you played it safe and stuck with the current node, you’d have been wrong & at a process disadvantage 14 times out of 14.

No improvement trend goes on forever. But why is this the moment lithium ion hits the wall? It’s like trying to call the end of a bull market.

2nm is just a commercial name. The real features are much bigger, see, for example, https://en.m.wikipedia.org/wiki/10_nm_process. So we are quite far from 10 atoms per transistor.

The biggest problem is lithography. ASML managed unexpectedly for many in the industry including Intel to solve technological problems with extreme ultraviolet sources, but shrinking transistors significantly further requires soft x-rays and the perspectives of that now for mass lithography are much more uncertain than for EUV ten years ago.

Not according to Jim Keller. He believes we still have plenty of room, and no offense to you, but I'll take his word before yours.

And Jim Keller is not a Fab / Processing guy. And if you actually listen to what he said, he is predicting plenty of room to improve, not it will improve every 24 months, which arguably has already stopped happening since ~2017

> So, this is about to stop. Very soon. Quantum tunneling, yield rates, etc are all starting to be cost prohibitive.

People have been saying that for over 10 years though. "Very soon" keeps being punted off by another 5 years every 5 years.

It has stopped, though. Single-threaded performance, either in absolute value, or per dollar has not been following Moore's law for more than a decade.

And outside of data centers, single-thread performance is still king.

Performance per watt has improved, but that's not a metric the typical end user cares much about.

Single-thread performance is not king though. Software nowadays is extensively multi-threaded. If you're still writing single threaded code for a desktop program people are going to look at you like you have antenna growing out of your head.

And performance per watt absolutely is important. People care a lot about battery life of their laptops and phones and that's strongly dependent on performance per watt.

Performance per watt is key for all portable devices, ie the laptops and phones that many do almost all of their end-user computing on.

We'll get to 256 core server chips and that will be the end of the line. You are forgetting that the advertised version number denoted by nm is about what size a theoretical planar transistor would need to have to be equivalent to the advertised process.

Since that number is purely theoretical, we can construct a theoretical scenario in which its theoretical nature would become absolutely obvious. Take a 7nm process node transistor and stack it 100 times. Such a process would be called a 0.07nm process.

> Even if hypothetically it does, there's a very real limit - an atom is only 10x smaller than 2nm.

That's certainly a real limit, but what is it a limit on? It is not obvious from first principles that it takes n atoms to implement n transistors.

On the one hand, I remember people saying the same about nodes much larger than those we currently use.

On the other, yes: continue current trends and single atom transistors become standard in 10-32 years.

I think it boils down to we will hit a wall, but we don't know exactly when we'll hit it. (and how hard, chances are that higher-hanging-fruit refinements will make the transition to stagnancy so gradual that we may not notice at all)

A jump from "all past predictions failed" to "and so will all future predictions" seems rather bold to me. In the end it's like a somewhat upended variation of the "x decades to practical fusion" thing where we all hope that the old joke that x might be a natural constant is eventually proven wrong.

Well, the main direction of refinements at this point seem to be around composable/heterogeneous computing where we basically have a lot of hardware optimized for specific workloads and throw the complexity at the software people. i.e. now deal with GPUs, DPUs, FPGAs, xCPUs, etc. instead of (largely) just a standardized set of CPU instructions.

Except we don't. Where's my single-core 100GHz processor?

Improvements still happen, but not always in the same way. If you implemented an application in 2003 assuming we'd have such a processor you'd be very disappointed. Counting on exponential improvements to continue is risky bet.

Scaling may take us to some strange places, but it's worth noting that an Apple M1 chip has >100x the transistors of a 2004-era Pentium 4, and achieves ~500x the FLOPs at a fraction of the power draw.

Yes this isn't single-threaded performance, but I think we should keep in mind that exponential improvement in price/performance over many decades is possible, if never certain.

Sources - https://www.alternatewars.com/BBOW/Computing/Computing_Power... - https://en.wikipedia.org/wiki/Apple_M1