Hacker News Reader

Comment by takoid | Hacker News Reader

M1 -> M2: 2301 -> 2598 = 12.91% increase

M2 -> M3: 2598 -> 3082 = 18.63% increase

M3 -> M4: 3082 -> 3864 = 25.37% increase

Assuming the M4 Geekbench score is legit, it’s interesting to see that the jump from M3 to M4 has the largest percentage increase in single-core performance since the M1 came out. The M1 -> M2 and M2 -> M3 upgrades were solid, but the 25.37% jump from M3 -> M4 is much more significant. It makes you wonder if we’ll keep seeing even bigger performance leaps with future releases, or if there’s something specific about this generation that’s driving such a large gain. Will Apple keep increasing performance at this rate, or is this an exception?

View on HN

Will they keep this up? Probably. Indefinitely? Likely not. But look at what it’s done. There are compelling ARM hardware outside of Apple now due to the renaissance in performance that this am series chip has shown.

I’ve an M3 max and now I can’t wait to see what an M4 max will do. 256GB ram in the top end laptop? 25-50% more cores?

> There are compelling ARM hardware outside of Apple now

Not only that, this may have finally forced Intel to innovate. Their new Lunar Lake CPUs coming out now are actually competitive with Apple chips for both performance and power. I thought I might not see X86 become competitive again in my lifetime.

> Their new Lunar Lake CPUs coming out now are actually competitive with Apple chips for both performance and power. I thought I might not see X86 become competitive again in my lifetime.

I hadn't heard about them, so Googled it and... "Manufacturered by: TSMC". So Intel's inefficiency was more due to their fabs being behind than inherent x86 downsides?

They might be competitive on battery-life, but definitely not performance per watt.

That's a good point.. I wonder how much network effect this has on substantiating ARM in server space too. Vendors have been trying that since the early twenty-teens and it was not going anywhere fast until Apple shipped the M1. I don't think it is a direct cause->correlation, but the existence of high powered dev machines running clang and gcc certainly lowers a lot of frictions even in higher level languages like Go.

AWS' Graviton line was released in 2018 to much fanfare, and there was a lot of content about all the cost savings per performance it brings immediately. Ampere were founded in 2017, and in early 2020 announced partnerships with Oracle for ARM based servers

M1 came out in November 2020.

A lot of Linux tools and systems and applications became available thanks to Raspberry Pis becoming very popular with hobbyists, and then Graviton pushing more "enterprise" stuff. For more user facing GUI-powered applications, it lagged, and still lags, but there Apple and Microsoft's ARM efforts were the main driving force.

Anecdotally, a lot of software in the graphics space didn’t have arm availability till after the M series were released. A lot of the windows/linux arm ports of libraries didn’t exist till after despite having hardware available earlier.

My work on Linux clusters was dramatically improved.

Having even one vendor prove that arm is worth paying attention to is the tide that lifts the ships of all other vendors in the arm space .

I wonder how much of this increase is solely enabled by TSMC process upgrades, either through higher clock speeds or increased number of transistors per core? It's kind of interesting that every iteration of the M series has corresponded to a change of TSMC process.

M1: 16 billion transistors at 3.2 GHz on TSMC N5

M2: 20 billion transistors at 3.7 GHz on TSMC N5P

M3: 25 billion transistors at 4.05 GHz on TSMC N3B

M4: 28 billion transistors at ?? GHz on TSMC N3E

N3E is slightly less efficient than N3B though, it just has higher yields.

Do both efficiency and yield play a part in the final off the shelf performance?

e.g. does higher yield mean you get more product in higher bins? I imagine we need to judge a process by the distribution across bins not just a max efficiency or a total yield.

Perhaps if this was intel or AMD with many binned SKUs, but Apple has very minimal SKUs and less binning wouldn’t account for as large a single core jump

Numbers are plausible. The M4 iPad is close to an i9-14900k in single thread.

I’ll buy an M4 mini when one comes out to replace my M1.

Same here.

I have been using an M1Max 14-inch for the last few years, but I seldom undock it, so an M4 Mini would be good. I don't really need the raw power of the Studio.

From what I hear, the next Mini could actually be "mini."

> The M4 iPad is close to an i9-14900k in single thread.

For how many microseconds ? /s

A good point which is why I want it in a Mac Mini where they (hopefully) haven't compromised it thermally as much!

Plausibly a large part of the perf bump comes from the SVE and SME extensions, for workloads that can put them to use.

> Will Apple keep increasing performance at this rate, or is this an exception?

I want to believe that but a lot of the performance gains have been the result of frequency boosting and you can not do that forever.

What is the current limit to that? Probably around 5Ghz right?

They are squeezing more out of the silicon but the frequency has been the primary contributor recently.

Those are nice numbers. But in reality Air M2 had worse performance, because cooling system was much worse.

I don't think Geekbench is a relevant benchmark for laptops. I would look at other benchmarks such as Cinebench, before drawing any conclusions.

Why not?

Geekbench usually doesn't run long enough to trigger the throttling on the laptop chips.

Reply on news.ycombinator.com