Hacker News Reader

Comment by bhouston | Hacker News Reader

Apple's custom ARM chips is definitely one of the things that does make Apple better. Low power, low/no noise, and high performance.

View on HN

Your sentence makes as much, arguably more sense if you erase the word "ARM".

the ARM chips are, to a large extent, what enable that.

Why does it matter what instruction set the chips support?

The main difference between x86 and ARM is that x86 is slightly harder to decode because instructions can be variable-width. But I have never heard of instruction decoding complexity being a particularly important bottleneck.

You've never had to rub a bag of frozen peas all over the bottom of your x86 Macbook Pro because it was overheating and you had an imminent zoom meeting you could not miss.

Those chips were designed and fabbed by Intel, not by Apple/TSMC respectively. That’s the relevant difference, not the instruction set.

The instruction set has only moderate impact on the chip’s frontend, and no impact on the backend. Most design decisions are unconstrained by the choice of instruction set.

The last few generations of x86 MacBooks were exceptionally bad implementations in this regard, and some of the better thermal behavior of the Apple Silicon MacBooks are things that they could just as easily done with an Intel CPU, if they had felt like it. For example, the Intel MacBooks was extremely eager to ramp its power consumption to the max, while the ARM MacBook slowly increases clock rate, one step at a time, such that it only hits max power after a long time of sustained demand.

I think that might have been Intels doing. They were on 14nm for 5 years or so, with each new 14nm release pushing the power budget and squeezing slightly more performance out of any corner they could find. I assume the CPU ramping was just another part of this approach. If the CPU ramps up faster it will seem fast to users and it’ll look faster in short benchmarks like Geekbench.

AMD doesn’t have that problem, though, so is it a problem with x86 or Intel? I would bet that Apple’s CPU team could get great results with a free hand on x86, too – probably not quite as good but close.

Memory ordering is also different.

ARMv8 was developed in order to make those chips. Something else could work but wouldn't be as good.

Are you suggesting that ARM isn't a large contributor to those things being possible? Genuinely curious.

The ISA is, according to people far more knowledgable than myself [0], not a significant factor in regards to performance and/or power efficiency on modern CPUs. That's also why both AMD and Intel haven't done a bad job keeping up with Apple on the efficiency front, AMD since the 4x00u series and Intel now with Luna Lake. Nowadays, the OS not having downright broken sleep [1] is what keeps Macbooks slightly ahead, though then again, MacOS makes up for that with other bugs.

Cause of that, if one is willing and able to go for a well supported Linux distro, they can in my experience get good battery life regardless of ISA. Essentially, whether on a Macbook with Asahi or a modern x86 notebook with either Linux or an aggressively managed/fixed Windows install, you'll do well. Personally lack the skill for the latter though, my Surface remains scolding hot whilst sleeping even after a fresh install, but c'est la vie...

[0] https://chipsandcheese.com/2021/07/13/arm-or-x86-isa-doesnt-...

[1] https://www.spacebar.news/windows-pc-sleep-broken/

What makes you think it is a large contributor?

Better nowadays. Previously Macs were widely laughed at as constantly having a generation or two (or more) older CPUs and GPUs, particularly for the price.

Thats not what I remember. Towards the end of the Intel era it really slowed down, but before that I felt that Apple was routinely releasing new macs with the latest Intel processors.

Would be happy to be proven wrong. Got some good examples?

High performance doing what? Everyone in high-performance computing is using Linux, my dude. Unless you're a studio who needs specific Mac workflows, or maybe you're doing LLM work?

Great perf-to-watt, but absolutely locked down so you can't do anything meaningful in the embedded space. I'm not over here mounting a max specced Mac Mini just to get that perf-to-watt.

We're using low TDP x86-64 when we need that.

building Ardour (.org) from source:

16 core Ryzen Threadripper 2950X: 9min

M3 macbook pro: 4min

Pretty meaningful when your life revolves around building Ardour.

Sure, but isn't that just comparing an obsolete CPU against a state-of-the-art CPU? I can't see attributing that to the ISA.

A scratch, optimized build of Ardour on a Ryzen 9 9950X takes 1m51s.

the machine is less than 4 years old!

anyway, the point was about the obsolesence or other of the ISA. It was about the claim that various versions of apple silicon are no good for performance, only for perf-per-watt or some related variance.

I agree with your second point, but misstating the age of a 6-year-old CPU as 4 years old will tend to throw off the equations, considering how fast this field moves. I would also note that the Zen/Zen+ CPUs were truly awful. Ryzen and EPYC did not have a good implementation until Zen 2.

2018 vs 2023 CPUs targetting entirely different segments.

Oh, I suspect they target exactly the same segments. The expectations of those segments have moved, however.

Reply on news.ycombinator.com