It's an ISA designed last century. It's time to stop tacking on extensions and build something better.
Even the time period is not the issue per se, though it plays a part in it. The real dividing line is more like 1985 than 2000 though.
CISC is the issue (at least, at the ISA level). Processors of that era were made to be "friendly" to assembly programmers. There are redundant instructions with slightly different semantics, there are many addressing modes, there are not many registers, and the instruction encoding is an illogical variable-length mess hidden away by the assembler.
RISC on the other hand starts from the premise that humans don't write assembly much anymore, they write high-level languages (yes even C counts) which get translated into machine code by a compiler. So an ISA is a compiler target first and foremost. So it's better to spend silicon on more registers and more concurrent/parallel work than on addressing modes and convenience instructions.
Obviously this battle has already been fought and won on a technical level by RISC; that's what the R in Arm originally stood for, and no modern CPU design even considers CISC. Heck, even x86 uses RISC-like micro-ops under the hood and the 64-bit version of it added a lot more registers. But the CISC instruction decoder is still there taking up some silicon.
The real question is: does that extra silicon actually matter that much in practice anymore? It's dwarfed by the other things a modern processor has, especially the cache. The answer seems to be contextual even today. An x86 server, desktop, or gaming console running at or near full load on CPU-intensive tasks still outperforms Arm (though by less and less each year), even though Arm handily takes the crown at lower power levels. Is that due to ISA alone or other factors too?
Except no one buys desktop computers any more: everyone's using phones, tablets, and laptop computers. There's also servers, but even here power efficiency is important; reduced power requirements for a datacenter would save a lot of money, not just in direct electricity consumption by the servers, but also from reduced cooling needs.
Wtf?
>Power efficiency
Isa has small impact on it.
A lot of regular office workplaces moved to laptops+docking stations as well.
People still buy desktops, but it’s a dying breed outside of gamers artists and those with heavy compute requirements
Outside of a couple unusual places like that, laptops have been the standard for workplaces over the last couple decades for me.
Before that laptops were not quite limited to managers, but they did signify higher status (if not necessarily pay). Today it‘s iPads that signify status.
Also what do you think most contributors to LLVM, GCC, CUDA, VFX,.. use as daily driver?
it's not clear that the 'contributors to LLVM' market is going to make up the shortfall.
https://www.statista.com/statistics/273495/global-shipments-...
ISA makes very little difference these days when it comes to power requirements.[1]
[1] https://chipsandcheese.com/2021/07/13/arm-or-x86-isa-doesnt-...
I think the lowest power ones are Atom and Celeron. All the other low wattage CPUs are RISC based. I think AMD is also coming up with something similar.
It's not as if there's something fundamental with x86 that makes it use more power. There used to be the case that x86 CPUs had many more transistors (and thus more power usage and heat) due to the need to implement more instructions, but I don't believe this is the case anymore. I think right now is that its easier to scale up an architecture, than it is to scale down.
If you look at something like Project Denver, it started as a x86 compatible CPU that would of had ARM-like power usage.
AMD Had Bobcat and later Puma, IIRC there were some <5W parts in both of those families. I still wonder -why- they never did a dual channel version of those parts for desktop, my guess is it would have made Bulldozer look extra bad somehow.