Thanks for your feedback. Let me see if I can understand all the points you raised.
- As far as I understand Ryzen has a pretty good branch predictor, based on a neural network which gives some speed gains. However I assume you cannot use it to skip the decoder? You need to turn an x86 instruction into microcode, no? Or do they cache and reuse previously decoded instructions? If so, that sounds interesting.
- One issue I see with your chip comparison is that AMD is making CPUs while Apple is making SoCs. One must assume a lot more chips on an SoC. Especially since they beefed up their GPU and Neural Engine as well.
About Power need and frequency. We know from Apple’s specs that their graphs for power usage are much more favorable than those for Intel. Yes power use increase with frequency, but the steepness of the curve, may not be the same for Apple Silicon and AMD Ryzen.
According to this Ryzen is using around 28W to maintain average clock speed above 3 Ghz. However there are 8 high performance cores, and you have only 4 of those on M1…
That is just a CPU though. M1 is an SoC with far more things on it drawing power. And one has to keep in mind that an AMD Ryzen needs about 5 GHz to beat a 3.2 GHz M1 Firestorm core. Thus had Ryzen tried to match Apple on power consumption they would not be able to match Apple on performance.
Ok, you can ask then why M1 Macbooks have amazing battery life? Secret is in energy efficient ‘icestorm’ cores that are significantly more energy efficient than ‘high performance’ cores in x86 processors. Obviously situation is reversed if we compare ‘firestorm’ cores.
I am sure that is a big part of it. But this is of course a choice that has been available to x86 makers for a long time and which they have not taken advantage of.
Apple has some interesting fresh ideas for CPU designs, but they didn’t reinvent the wheel.
I hope I haven’t given that impression. As far as I am concerned this is not about inventing new ideas but leveraging clear strategic advantages that Apple enjoys:
- Using an ARM ISA allows them to double the number of decoders. Not a simple choice available to x86 chip makers. Is that a genius on Apple’s part? Nope, they are just exploiting an opportunity given to them by using an ARM ISA.
- Aggressive adoption of heterogenous computing. Is that some stroke of genius or insight from Apple? Nope, they are just exploiting the fact that they have full vertical integration and can easily go down this route.
Majority of m1 advantages relies on significantly better manufacturing process.
That is a totally unfair assessment I think. Apple has advanced the performance of their chips far more rapidly than other ARM makers and Intel. And that cannot be brushed aside simply by saying: “Oh it is nothing to see here, just manufacturing boys.”
Manufacturing is part of the story, but claiming it is the whole story is very uncharitable towards the work Apple has done.
I will be really surprised if Zen4 based on 5nm wont blow away Apple Silicon.
Hahaha the game is on. We need to do some kind of bet on this. Not on money but some kind of public acknowledgement that the other guy was right. I have to think about this first though. Whomever release a chip last will have an advantage there and then.
Hence just when Zen4 is released it may very well be faster. But what happens when a M2, M3 or whatever is released afterwards. The question is if AMD will be able to retain a performance lead very long.
By the time AMD gets on 5nm, Apple will be on 3nm, so they will get more cores than them anyway.
It doesnt even matter what Apple do — 16 high performance cores on M1X? Its not problem — in chip on similar die size AMD will be able to pack 20–24 cores.
Here is the problem: You keep comparing a CPU with an SoC. Also Apple is pursing heterogenous computing solution. In fact I believe there has been comments about how certain parts of the Firestorm cores are weaker on particular computations because Apple knows their co-processors/accelerators will take care of that.
This makes a comparison with Zen4 and Firestorm a bit of an Apples to Oranges comparison. Apple is seeking maximum real world application performance. A lot of the most performance intensive tasks will will not be running on the Firestorm cores, but on Apples GPU, video encoding ASICs, as well as Neural Engine. That is why e.g. an iPhone has long been able to outperform high end Macs on stuff like video encoding and playback.
That is magic of x86 that Apple can’t replicate
There is not magic in x86. Quite the contrary, it offers nothing but disadvantages. It is only through heroic efforts and throwing money at the problem that AMD and Intel have managed to keep x86 ahead of the competition. In terms of squeezing out performance the x86 ISA is a major obstacle. Code density is weak, which was supposed to be a CISC advantage. They cannot throw in more decoders due to variable instruction length.
The number of registers is also still a disadvantage. x86–64 gives just 16 general purpose registers. ARM A64 has 32, twice as many. x86 has 16 SSE registers. Again ARM has twice as many.
There isn’t a single trick that x86 can pull which an ARM CPU cannot do as well. But there are lots of advantages that ARM has which x86 cannot replicate such has having fixed sized instructions and twice as many registers.
complex and high transistor density of M1 also has many disadvantages.
It is an SoC so of course it takes more transistors, but that also gives big advantages such as high speed unified memory, and application specific silicon to accelerate demand tasks. That translate into real world performance benefits for actual applications people use.
Excellent relations with TSMC will be always biggest advantage of Apple Silicon.
That is not a random transient thing. Apple has 60% of the profit in the PC market and something like 90% of the profit in the mobile market. That means Apple is able to pay the fixed cost for high density chips. Making a 5nm or 3nm chip is really expensive. If you don’t have enough volume and profit from your sales you cannot do it.
A lot of companies will not be able to follow Apple into 3nm because they don’t have Apple’s volume and profit. Even AMD may struggle to do that. That isn’t Apple magic. That is simple Apple having a business model that allows that.
They sell premium products which allow them to command higher margins and get more profits.