They have certainly built an amazing architecture.

I'll say that I stated it rather too miserly, a bit out of annoyance from the extreme hyperbole surrounding the M-series processors ("its better than an i9 + 3080 system! Next year it'll beat an i9 + 4090!!").

Funnily enough I plan to switch to an M-series Macbook as soon as Asahi Linux is completely out of beta.

granted, I think maybe more eyes are on this snapdragon release as it might have some Apple/Nuvia special sauce in this batch. Albeit not sure how motivated Gerald Williams and co are to be designing mobile/laptop chips again given the whole point of their leaving were to do server chips...

You are a fool for believing the 800Gb/s memory bandwidth meme that is only true for the most expensive chip. The chips most people can buy are nothing special and actually quite underwhelming when you consider the price. GPU/CPU integration is also a meme that keeps on giving zero results. Supposedly so much better yet everything runs slower than the competition. It consumes less power? It better with node advantage bought with monopoly money it would be a shame to both suck and use more ressource. And then you pretend they make any progress at all when the truth is when you account for node upgrades, frequency boost and similar upgrades they are not making any progress at all. But I guess double standard is a de facto status of the Apple fanboy.

Apple is doing something interesting but it's not at all unclear how they're doing it:

> outpace on a per watt basis with AMD/Intel (and be at least competitive with on an absolute basis)

They're the first on TSMC's new process nodes, so whenever a new node comes out, people will compare the new Apple chip on a new node to the previous generation AMD chip on the previous node. If you compare them this way then the newer node outperforms the older one as expected. Then AMD releases one on the new node and the advantage basically disappears, or is just making different trade offs, e.g. selling chips with more cores which consequently have a higher TDP in exchange for higher multi-thread performance.

Intel hasn't been competitive with either of them on power consumption for some time because Intel's fabrication is less power efficient than TSMC's.

But if Apple's advantage is just outbidding everyone at TSMC, that only lasts until TSMC builds more fabs (the lead times for which get shorter as the COVID issues subside), or someone else makes a better process.

> For example, they have a unified 800GB/s memory bandwidth with zero copy sharing between GPU and CPU. That’s not something anyone else has managed yet.

It's also not something which is at all difficult to do. It's a straightforward combination of two known technologies. Integrated GPUs have unified memory and discrete GPUs have high memory bandwidth.

There is no secret to attaching high bandwidth memory to a CPU with an integrated GPU, it's just a trade off that traditionally isn't worth it. CPUs typically have more memory than GPUs but if it's unified and you want to use the fast memory you're either going to get less of it or pay more, and CPU applications that actually benefit from that amount of memory bandwidth are uncommon.

One of the rare applications that do benefit from it are LLMs, and it's possible that's enough to create market demand, but there's no real question of if they can figure out how to make that -- everybody knows how. It's only a question of if customers want to pay for it.

And what we may see is something better -- high bandwidth memory as a unified L4 cache. So then your CPU+iGPU gets the amount of HBM traditionally found on a discrete GPU and the amount of DRAM traditionally attached to a CPU and you get the best of both worlds with no increase in cost over a CPU + discrete GPU. But currently nobody offers that -- the closest is the Xeon Max which has HBM but no integrated GPU.

And none of these are why Qualcomm isn't competitive with Apple -- they're not competitive with AMD or Intel either. It's not the fabs or the trade offs. Their designs just aren't as good. But all that means is they should hire more/better engineers.