Comment by cogman10
1 year ago
Or CPUs really. Die shrinks just aren't giving the advantages they once used to.
You can see this in the fact that RISC-V and ARM architectures have caught up with x86 performance even though x86 has had a decade long head start and billions more invested in development.
We are quickly approaching a weird space. Barring some major innovations, you are likely to see that 10 year old equipment remains competitive with brand new products in terms of performance.
ARM has gotten very good, and is definitely competitive with mid-range x86 while offering better performance-per-watt, but it is still not competitive with high-end x86.
Depends on the task and what you're measuring.
My M1 (standard) CPU is definitely faster than my Ryzen 7700x on some tasks, and it blows it away on perf/watt.
Those are both mid-range CPUs in standard consumer configurations, and I agree ARM does very well in that segment, and it does even better in the low-power/mobile/embedded segment where x86 is practically non-existent (recent gaming handhelds notwithstanding).
However, high-end workstations, compute-focused servers, and supercomputers, which use extremely expensive and power-hungry x86 chips, are a different segment, one to which ARM currently has no direct answer (and some might argue it shouldn't have one because such wasteful things shouldn't exist). This segment once had a number of competitive RISC players, like POWER and SPARC, so I don't think it's unobtainable for ARM.
2 replies →
> You can see this in the fact that RISC-V and ARM architectures have caught up with x86 performance
Concerning RISC-V having caught up with x86 performance: dream on ... :-(
this was pretty impressive https://news.ycombinator.com/item?id=41364549
What do die shrinks have to do with ISA performance? Also, there are no RISC-V CPUs available that match the latest X86 or ARM CPUs. Even then, the ISA chosen doesn't have much to do with the performance of CPUs (at least, when comparing major ISAs like X86, ARM and RISC-V).
Smaller nodes (generally) translate to lower power consumption, more transistor density, and faster transistor switching speeds.
It allows for an architecture to deepen pipelines, add registers, add cache, and pull off tricks like SMT.
> the ISA chosen doesn't have much to do with the performance of CPUs
That's somewhat my point, the limiting factor for ISA performance is the physics surrounding the transistors.
Yeah I keep looking into upgrading my 12 year old PC, but for like £1500 I can get one 10x faster (multithreaded) and only about 4x faster single core. I mean, that's a decent boost but it feels very disappointing for 12 years of progress.
That might've been a result of Intel having the best leading-edge fabs until 2018 or so. It was hard to judge different ISAs before then.