Its time to design a new Instruction set architecture | Hacker News Reader

Its time to design a new Instruction set architecture

news.quelsolaar.com

5 pointsquelsolaar7 comments

> It would however surprise me greatly if Apple makes the investments needed to make their architecture competitive on the high end, given how small they are in that market

Does anyone know what is meant by this?

My understanding is that while not the absolute best, Apple’s cores are already competitive at the high end.

Do they mean Apple likely won’t build high throughput server chips?

The development of Thread-ripper and Xeon product lines are amortized over a huge amount of servers/workstations from many vendors. It makes no sense for Apple to spend that much to develop a competing CPU for the tiny fraction of workstations that Apple makes. For Intel and high-end (and gaming)is where they make their money and margin so thats a big focus. Apples high-end workstations are a rounding error for Apple in terms of revenue.

Sure - understood - that’s essentially what I just said.

But looking at cores themselves, Apple is already nearly competitive with AMD, and ahead of Intel.

My understanding is that the things which still differentiate these two ‘high end’ example’s from Apple’s are essentially down to core count, and cache.

What I’m asking is whether there are other architectural features that you have in mind that I’m overlooking that Apple has not developed, but which AMD and Intel have?

If you are mainly focused on power consumption as Apple is you dont want to add more logic to squeeze out the last bit of performance the way Intel and AMD does. The last 20% is a lot more expensive than the first 80%. It its amazing that AMD and Intel are beating Apple (even if its not by very much) given what they have to work with. Its takes A LOT, to make x86 to run fast.

AMD and Intel, doesn't tell me about their secret sauce (And I'm under NDA) :-( I know they have crazy good branch prediction that apparently uses machine learning, and a lot of other tricks that takes a lot of gates to make work. If they could leverage that technology on a better ISA they would get a lot out of it.

If high performance is what you want, you move to a chiplet design so that you get higher yield, you have a separate GPU, you add more IO and things like that. I think going forward, that Apple will continue to take design decisions that favor low power consumption, and that means they wont be able to scale that design as well to the high end.

Ok, but as far as I can see M1 cores aren’t 20% slower than Intel or AMD cores. They are faster than Intel cores, and in the same ballpark although in aggregate slightly slower than AMD threadripper.

This would seem to negate your analysis, if you are simply talking about the trade off between power consumption and performance.

Also you seem to ignore that higher performance per watt means you can get higher absolute performance before you hit the thermal ceiling.

The entire point of the article is that AMD/Intel aren't as fast as they should be, give that they are using such an out dated ISA, that's why they should develop a new one.

Yeah - that’s not a crazy conclusion.

I just don’t find the way you get there to be very clear.

Reply on news.ycombinator.com