I'm not sure focusing on the instruction set is the right angle.

x86 would be more competitive in the mobile space if Intel had offered efficient chips. Is the instruction set the biggest blocker that prevented this? Or is it the licensing model? Perhaps if Intel had licensed their architecture and instruction set to others, like ARM does, they would be in a better position?

ARM's licensing model get's a lot more eyes on instructions sets and chip designs. It allows big companies like Apple come in and run the show, licensing just the instruction set. Why wasn't Intel working with large customers with an interest in custom x86 designs? In hindsight it seems presumptuous to assume there is no x86 talent outside Intel/AMD.

Really impossible to know how much the actual instruction set is holding Intel back without being on the inside.

The problem is that the revenues from those licenses would be so much lower than the margins on those x86 CPUs to make it almost impossible for Intel to switch. Apple moving Macs to Arm has given better performance __and__ lower prices.

I do think that Intel has executed badly on the ISA front too. Let’s not forget that x64 wasn’t an Intel design (how much did they spend on Itanium) and they’ve messed around with AVX-512 which just seems to eat die area and power.

By contrast Arm managed to get its customers switched over to 64 bit seamlessly and has left most of their awkward legacy behind.

true. but I really cant imagine that the area and latency associated with the uop translator is really where this is all falling down. i have always wondered how much support for the old vector units, segmentation and all the rest cost, but it cant really be that much since caches keep getting bigger and bigger. its a really good question.

There are other costs that either make very wide OoO more difficult or more costly. x86 has a lot more flag-based instructions compared to Arm. That adds more dependencies that the reorder engine has to sort through. x86 variable length decoding takes log(n) in decode circuit depth, which either forces a longer pipeline or limits clocks. And obviously AVX512 units are just huge because a decision was made to make them the same latency as normal MUL/ADD/FMA. And x86 designs have to scale in clockspeed from server / tablet (~2.5GHz) to desktop and high perf laptop (5GHz+). That forces suboptimal designs like the 5 cycle L1 in Golden Cove. Meanwhile Apple has a 3 cycle 192kB L1.

Yeah I think this is all valid - i really mean it more as a product line and the lack of practical substitutes before recently (especially on the server side), not the specific details of the instruction set v others. They were basically the only game in town for a while. Had they licensed it, yep, it would have been different.

It’s a classic innovator’s dilemma- they can continue milking the x86 or reinvent the company but not both at the same time.

> but not both at the same time.

For 20 years they could have afforded to. Now it may be too late. If they try I don’t think the market would punish them for it.

And in fact they are famous for killing their cash cow (memory) and focusing on the new thing (microprocessors) in the 80s under Grove. They have fallen far.

Other big companies, like Google, have this problem and haven’t made efforts to address it.

Intel is not going to beat anyone else in architecture or design in any compute system besides x86. They don't have the experience, the talent, or the culture to do so. Heck, Intel can't even beat AMD in x86 power/perf. So Pat is just stating the obvious, Intel has no option but to go maximal x86.

Yep, I'm not even saying he's making a mistake. Every option looks problematic and his choice might well be the least bad option.

If Microsoft managed to do turn around, then why they cant?

The specifics of the situation are very different.

What do you mean?