While you are debating if eternal growth exists we that actually write software can tell you that the bottle-neck has been RAM since 386 that introduced caches:
DDR3 was the peak of RAM latency, 65nm SSDs where the peak of flash memory longevity both in ~2010!
You cannot scale anything unless it's embarrasingly parallelizable in which case you never had a problem in the first place.
Peak performance is more important than the puny 0.5Gflops/W "5"nm M1 (2.5Gflops/W) has over the 28nm Raspberry 4 (2 Gflops/W).
Everything that is below one order of magnitude improvement is irrelevant = ASML is irrelevant unless devices break and you buy things you cannot repair.
That combined with eternally raising electricity cost, not as a function of marginal prices on a market (like today), but as a permanent depletion of resources leads to only 1 conclusion:
Don't wait for the next generation hardware/firmware locked, rentseeking line of hardware: 14nm X86 and 28nm ARM IS good enough for a lifetime if you control it and make sure it lasts.
There are only 2 programming languages that can survive this long term: C on the client (with a little ++ for convenience) and vanilla JavaSE on the server (with as little deps. as possible).
Bonus prediction: JavaME will be revived in some form for open microcontroller mobile use.
> Peak performance is more important than the puny 0.5Gflops/W "5"nm M1 (2.5Gflops/W) has over the 28nm Raspberry 4 (2 Gflops/W).
Those numbers are very misleading - perf per watt is not a constant but something that very much depends on how far the clock speed is pushed. Anything tuned for perf is going to have significantly worse perf per watt than the architecture is capable of. The same goes for the CPU design - maximizing perf and maximizing perf per watt is going to give you different architectures on the same node.
DDR3 was the peak of RAM latency, 65nm SSDs where the peak of flash memory longevity both in ~2010!
You cannot scale anything unless it's embarrasingly parallelizable in which case you never had a problem in the first place.
Peak performance is more important than the puny 0.5Gflops/W "5"nm M1 (2.5Gflops/W) has over the 28nm Raspberry 4 (2 Gflops/W).
Everything that is below one order of magnitude improvement is irrelevant = ASML is irrelevant unless devices break and you buy things you cannot repair.
That combined with eternally raising electricity cost, not as a function of marginal prices on a market (like today), but as a permanent depletion of resources leads to only 1 conclusion:
Don't wait for the next generation hardware/firmware locked, rentseeking line of hardware: 14nm X86 and 28nm ARM IS good enough for a lifetime if you control it and make sure it lasts.
There are only 2 programming languages that can survive this long term: C on the client (with a little ++ for convenience) and vanilla JavaSE on the server (with as little deps. as possible).
Bonus prediction: JavaME will be revived in some form for open microcontroller mobile use.