Live data from Hacker News

M3 Macs: there's more to performance than counting cores

eclecticlight.co

371–380 of 419 posts

Re: M3 Macs: there's more to performance than counting cores

#371
I asked Claude.ai to summarise this discussion for me, and below is its expanded summary of the discussions related to Apple Silicon and the M1/M2/M3 chip generations.

ON M3 PERFORMANCE GAINS:

The jump from M2 to M3 seems smaller than M1 to M2, especially for the regular M3 chip. Some speculate this shows limits in further scaling down process nodes.

The M3 Pro reduced performance cores from 8 to 6 compared to M2 Pro, suggesting a focus on efficiency, yields and costs rather than maximum performance.

But the M3 Max increased performance cores from 8 to 12 versus M2 Max. It matches or exceeds the M2 Ultra, indicating significant gains at the high end.

The M3 line overall shows Apple is diversifying core configurations across the range. This makes comparing performance gains more complex than simple core counts.

There is debate around how much benchmarks like Geekbench reflect real-world usage. Things like frequency management and cluster configurations affect multicore performance.

The 20-30% gains over M2 may seem small yearly, but comparing to M1 shows solid doubling of performance over 2 years.

ON EFFICIENCY CORES:

Some argue the M3 wasted the process shrink on just frequency gains rather than wider cores or gains in other areas.

But others note efficiency is still improved. The regular M3 efficiency cores clock higher while using the same power as M1's.

Efficiency cores handle lightweight threads well, not just background tasks. Their performance can be significant when frequencies are maximized.

ON THERMAL MANAGEMENT AND BATTERY LIFE:

The M3 line generally favors efficiency and thermal management over peak performance. This suits laptop use cases well.

But it's unclear if a future M4 will tune for performance over efficiency in desktop models, or just add more cores.

Either way, the 22 hour battery life on M3 Air is seen as a major achievement showing efficiency strides.

Re: M3 Macs: there's more to performance than counting cores

#372
post #15

It seems like the M3 favors efficiency and thermals over raw power. This makes sense for the devices they've put it in. I wonder if we're going to see different processors (M3X or M4) for the next release of the pro desktops, that favor power over thermals and efficiency? Maybe there will be a kind of tick-tock, with odd numbered processors favoring efficiency and even numbered processors for power? M3, M5, M7 for th…

Apple is favoring yields and margin over everything else. They cut the M3 Pro as much as possible so it's half the area of the M2 Pro.

That would explain it. Reduced chip size, presumably due to increase in price per transistor. But do you have evidence for this theory?

Re: M3 Macs: there's more to performance than counting cores

#373

Earlier quoted context omitted.

Why are you arguing about PC performance with your friends? Don't you have something better to do? The biggest advantage of ARM over x86 is security, efficiency is second. Performance is nice to have but is kind of coincidental (depending on workload) and I'd expect everyone else to catch up.

Are you lost? This is HN, I talk with people about computers a lot.

Yeah but you probably shouldn't. In my experience enthusiasts/power users aren't actually any more accurate on this subject than, say, your grandma. They just know bigger words they can use in conversations.

So there's got to be something either more productive or more grass-touching to do.

Re: M3 Macs: there's more to performance than counting cores

#375

Earlier quoted context omitted.

Hardly any AAA companies are optimising for the Apple GPU. MoltenVK is where all the interest is. Even if they did, there is very little in Apple Arcade which taxes the GPU, most target the lowest common denominator in terms of supported iOS/phone combinations. The original Apple Arcade strategy was for AAA titles, but for whatever reason that wasn’t pursued, so now we have a tonne of casual games and re-releases of…

> The original Apple Arcade strategy was for AAA titles Can you prove that statement? I bet you can't, cause it's simply not true. Every Arcade title is playable from an iPhone to a Mac by way of Apple TV. It was never going to get AAA titles.

Here you go. A simple google would have found that.

Apple wants more Grindstones.

https://9to5mac.com/2020/06/30/apple-arcade-strategy-shift/

Re: M3 Macs: there's more to performance than counting cores

#376
post #324

Earlier quoted context omitted.

I might argue that what you are describing is exactly why it's a bit absurd. It would be surprising to me as a user to discover such a limitation. If I didn't read about it here, I would probably find out after the return window and then be pissed that they took a stand on this particular artificial market segmentation.

The vast majority of people wouldn't return something because it doesn't have a feature they weren't even aware in the first place that it didn't have, because they don't need it. That's just weird.

Perhaps they should also only allow users to plug in an external mouse if you buy the pro or Max chip.

Re: M3 Macs: there's more to performance than counting cores

#377

Earlier quoted context omitted.

> The vast majority don’t care or even own an external display. Or more than 1 external display. The vast majority own none or a single. Few people have multiples, and those people are generally enthusiasts who will also tend to pay more.

I'm curious where people are getting this data point?

Fair point. It's a bit anecdotal and things I've read over the years. Multi-monitor was growing quickly for desktops until the big wide screens came out. Then it seemed to slow. As many people shifted to laptops, the one larger monitor seemed to be more common or no monitor at all. There's also the challenge of defining multi-monitor as I use 3, but one of those is my laptop. In that sense nearly everyone can run 2 screens.

I'd love to see some modern stats that take into account these nuances.

Re: M3 Macs: there's more to performance than counting cores

#378
post #184

Earlier quoted context omitted.

> A good bit of boils down to macOS expecting monitors to be well-behaved, e.g. each having unique EDIDs (many don’t, instead sharing one across all units of a particular model) and initializing in a timely fashion. I don't really buy this explanation. I'm talking about using a single external monitor here, a monitor which I regularly also use with linux (intel/amd) laptops without a single issue. > Because it’s full…

> I don't really buy this explanation. I'm talking about using a single external monitor here, a monitor which I regularly also use with linux (intel/amd) laptops without a single issue. Might be model-specific then. It’s not something I’ve seen with displays from Asus, Alienware, and Apple. I briefly owned a Dell monitor that would periodically flicker but it got returned. > That's not really a good reason though. I…

In my experience, MacOS monitor support is very hit or miss. For an Asus Gaming monitor, I had to write a custom plist file to get it to even show a picture and getting it to run at native resolution took a lot of experimenting.

I would not buy any new monitor for MacOS that does not explicitely advertize MacOS support.

Re: M3 Macs: there's more to performance than counting cores

#379

It’s ridiculous that the base and Pro only support 1 and 2 external displays, respectively. Want 3 monitors? Have to pay for the Max. https://www.macrumors.com/2023/11/02/m3-chip-still-supports-...

It’s called command right arrow. Try it. I think someone did a study and found people cannot use more than two. Everything else you’re better off using desktop management.

Re: M3 Macs: there's more to performance than counting cores

#380

Earlier quoted context omitted.

I was thinking about the desktop variants. Apple Silicon is amazing. But if you’re doing heavy GPU it’s not very exciting.

It's very exciting in the local LLMs space, where the unified memory allows fast inference of large models.

I read that the bandwidth is 50% lower on the M3 combined with a lower CPU core coun. This reduction will impact inferrence performance. It maybe better to stick with M2 series if you really really think spending $10k on a laptop mac to do inferrence slowly (but faster than a plain PC of course) make sense.
Post reply on HN