The ARMy is invading

Intel did try moving away from 14nm, but their 10nm efforts went completely bust and then they had to scramble to get 7nm working in time to not fall too far behind AMD anymore.

Anyway, it doesn't matter, because I imagine even an average ARM CPU will probably still be a better pick for mobile devices than the most efficient x86 CPU. The big interrogation mark is software compatibility. Chrome books didn't really catch on, but then maybe people really wanted Windows.

Indeed! Our work machines were recently upgraded to M3 MacBook Pros with 64GB (genuinely necessary for the work we do) and it really does feel like an impossible amount of power for a unit that I can work really hard for 8 hours on one charge and literally never hear it. It doesn’t even get particularly warm.

Game changer in every sense. And for the Studio boxes they staple two of those fuckers together in “Ultra” configuration. It’s scary good. And, Apple being Apple, insultingly priced.

Yeah, I hate Apple generally, but the M chips are really good. Absolute pain in the ass programming on them a couple of years ago whilst software devs had to fix all their packages, but now it's mostly smooth sailing.

Eh it's somewhat illusionary. Most stuff except x86 has been RISC since the 80s. PowerPC chips (as seen in old.macbooks, the 360 and PS3) were RISC, ARM was originally designed for the BBC Micro in effect. So it's not even remotely new.

They've always been more efficient. Which hypothetically means if you clocked one high enough they always had a raw power advantage over CISC. Less power per instruction means if you ram the same power through it you do more instructions.

The main changes are economic. Amazon and Apple have gone in designing their own ARM chips, as have Nvidia with TEGRA. Where as previously the R&D for them was going into making them lower power for mobiles as that was the way for Qualcomm etc to grow. Now there's tech giants trying to get the performance out of them.

That all said since Core 1 decades ago, Intel chips internally run something that's basically RISC. The pipeline is heavily based on RISC principals and nothing like the 40 stage pipeline in the Pentium 4 which was the last true CISC design. Internally the first stage of a modern intel chip is a decode layer into an internal RISC system. Intel maintained a genuine lead for most this time because they added more "features" for peak performance. Internal GPUs, SIMD extensions, hyper-threading etc. the new hotness being AI co processors. The main change of late has been Apple, Amazon, Google actually adding those features to ARM (or paying ARM to do the R&D).

Back when I were a student the new hotness R&D from ARM was clockless asynchronous processors.

Yeah, here I’m complaining that windows will be taking screenshots, but meanwhile Apple knows everything in my photo library.

I think that's very different for a camera app to look at stuff that's in the viewfinder when you press a button to just slurping everything that appears on your screen, and it's the kind of difference AI tech bro types will try and make seem like it's less than it is in order to try and make this kind of shit a fait accompli.

It's part of the reason the hype artists have been pushing for grouping previously existing ML-driven processes like computational photography alongside the theft of work at a massive scale that is things like Midjourney and the various LLMs that have sprung up under the banner 'AI'.

Mate, if they wanna put close-up pics of my oozing bangle into one of their image models, more power to them.

The first super scalar CPU was in the 60s and the first common implementation was the Pentium 1 in the mid 90s.

RISC designs like ARM fundamentally date to the 80s. Silicon changes slowly.

Plus the neural engine and tensor chips aren't that specialist over a GPU. Or they weren't originally.

But going to heterogeneous computing with SoCs is more of an economics change regardless. It's been in progress for 15 years or so.

I feel like we're finally returning to something more like the 90s. When I was a student (in the 00s) learning all this stuff there were all sorts of major people making chips, Sun with SPARC, IBM with PowerPC, ARM, Motorolas plus all the mad shit like custom Cray CPUs. Though that was a reduction from the glory days.

By the 2010s we hit an absolutely dreary period of it being an x86 or an ARM (but only in mobile devices). This is getting back to the fun times of co-processors for specific jobs. Though I'm not sure the major steps are all that shortened, they're by and large incremental.

Like when I did my Phd everyone was adamant that the MESI protocol for cache coherence could not scale past 16 cores on a die. Low and behold that basically still holds true, Intel are only edging around it because of the "efficiency" core thing. They've just got a shit load more cache on chip these days. 

Feels kinda fun. But also that people have been saying this shit for decades, just Intel's dominance kinda kept it under wraps.

I wonder if I should invest in some ARM shares seeing as how Nvidia are seemingly the world's most valuable company.

Everyone's favourite grandpa investor, Buffet, always recommends you invest in what you know. I guess I need to purchase a load of shares in mustard.

Do Qualcomm have their own fabrication plants? Intel's are so big they can't even fill the manufacturing capacity with demand for their own chips, and there was on/off talk about spinning the manufacturing bit off entirely.