Comment by cwizou | Hacker News Reader

Comment by cwizou | Hacker News Reader

Parent Full thread

While I could agree with the general sentiment, I think it's hard to understate how much of a role Apple played in the background of all of this.

But in any case, there's plenty of things to be said about this article. About one year ago (random link with relevant quotes : https://www.pcgamer.com/intels-3d-chip-tech-is-perfect-so-it... ), Intel was mocking AMD for using a chiplet approach, before announcing today that it was - clickbaity title aside - going to change everything.

The sad truth is, both Intel and AMD are in the exact situation. AMD went chiplet in order to make their performance cores at TSMC, and their less critical cores ("IO") at GloFo.

And Intel will be doing the same thing tomorrow (again, random link on the topic: https://www.tomshardware.com/news/intel-ceo-visits-tsmc-agai... ) by producing their performance cores at TSMC and their less critical ones on their own processes.

In both cases, this is just a question of using a very limited resource (TSMC's best in class process) the more effectively that you can (by throwing extra engineering at making a chiplet design that works).

And it's supremely relevant to the discussion to talk about how Apple, by throwing capital at a company (TSMC) that was, for the couple of decades I used to cover this, at best 2 years behind the best in class, today where they are (far far in front).

We could definitely have a long discussion about the hubris that led 2015 Intel where they are today (completely stuck with a 7 year old "+ paint coatings" aging 14nm "performance" process), or how Gelsinger is trying to make the best out of the situation (I personally think he's immensely qualified and Intel's best hope, though that may not be enough to bring Intel back to where it was), but at the end of the day, Apple threw a wrench in what seemed like an unshakable performance lead from Intel by spewing a bit of money left and right (they didn't only bet on TSMC early on, they threw money at GF for example, and it wasn't that massive early on from my understanding), and the silicon world hasn't been the same since.

>a role Apple played in the background of all of this

All of what? This topic is about Intel chilpets, which many of those will end up in datacenters where most Intel chips go, and that's not where Apple sells chips for.

Not every chip made and sold in the world revolves around laptops, tablets, smartphones or the apple ecosystem.

So can we please talk about Intel's chiplets impact on the industry and less about Apple silicone which has nothing to do with this?

Maybe I wasn't clear or went too fast on some things, I'm not a native speaker.

This is all about semi manufacturing, and the position that TSMC now has in the fab space, thanks to Apple (yes, really, that's what I went on about in the previous comment). No part of my comment referred to arm, architectures, anything of the sort, just that Apple's money, applied broadly at first in the semi manufacturing space, then in a very very targeted way, took TSMC from tier 2 manufaturer to the best in class.

If you look back a few years, only x86 chips were having volume at the bleeding edge of manufacturing process. Gpus, Smartphone, everything else was one node back at the very least. Apple threw money and orders with a massive volume (iPhone + iPad is pretty close in units to the x86 cpu market, above 300M roughly off the top of my heard) at TSMC and that early + continous investment helped them fast forward their processes while Intel is still stalled in 2015.

Apple is using TSMC today (the best bits), AMD is using TSMC today (the second best bits) for the performance part of their chiplets and so will Intel tomorrow for the exact same reason. This is the relevant bit that I was pointing at.

That's a fair assessment although perhaps not giving AMD enough credit.

IMO GloFo's spin-off worked out very poorly for AMD in the short term, but long term it let them partially-leapfrog Intel much as they had done 20 years prior with the K6/K7.

There's two main things that IMO give the M1/M2 their 'magic';

- Dram on die (helping their PPW, especially single threaded PPW) - Tight integration between OS and CPU.

This is, perhaps, where the x86 consortium has fallen into a challenge in the face of tight integration; The majority of that group would likely shriek at the idea of a DRAM on CPU, "here you go that's all you get" idea. I saw it a lot when I slung PC hardware; Folks who would insist on having as much upgrade-ability as possible, but never actually bought the upgrades between PC purchases. Even still, RAM is the main thing I personally find myself still upgrading on either purchased or older PCs.

That being said, It would be interesting to see if they try doing DRAM chiplets for these; I'm sure some 'ideal state' would be where DRAM chiplets + slotted RAM cause the chiplets to be dedicated to integrated GPU resources, or act as a form of L4 cache for one or more banks of DRAM.

> This is, perhaps, where the x86 consortium has fallen into a challenge in the face of tight integration; The majority of that group would likely shriek at the idea of a DRAM on CPU, "here you go that's all you get" idea. I saw it a lot when I slung PC hardware; Folks who would insist on having as much upgrade-ability as possible, but never actually bought the upgrades between PC purchases. Even still, RAM is the main thing I personally find myself still upgrading on either purchased or older PCs.

Memory people usually either buy as much as they need, or buy some and then add more. I've done it just about every PC build of mine, sometimes completely swap memory.

GPU generally gets upgraded, unless you're one of those who buy the best every year.

Storage for sure gets upgraded.

Currently sending this from my skylake i7 that seen 3 different GPUs, 2 different RAM kits, I lost count how many times I've upgraded storage.

I would probably get mad if I couldn't upgrade memory down the line. It maybe makes sense for laptops, but I don't see why do that on desktop. While M1 has memory right there, its latency is higher than intel and amd.

Reply on news.ycombinator.com