|
|
|
|
|
by zamadatix
2031 days ago
|
|
The ISA is open but the chip is custom and very different from any of the standard designs ARM Holdings has (or anybody else for that matter). Not exactly an "architecture" to open like OP said but just pointing out the CPU isn't some standard design with nothing to open up simply because it licenses the ARM ISA. That's like saying your book is off the shelf because it used an English dictionary for the words. The vertical software integration creates many small wins that aren't insignificant overall but it's not why they are nearly 2x ahead of Qualcomm using the same ISA at the same power level in 3rd party userspace computation only benchmarks. There isn't any vertical integration in those, the raw performance is just better. For another take on this Chrome vs Safari benchmarks put Chrome at ~90% the speed in web benchmarks and this is Chrome's 1st stable release of an M1 native version. At the same that's 45% faster than Chrome on my AMD 3950X. I remember when Apple added the Javascript rounding instruction to their chips people speculated that's why Safari was now so fast, turns out a Safari dev chimed it they hadn't started making use of it yet the new chips were just faster than before. |
|
The implementation is not an ARM reference design; but it's Apple's own implementation. Why should they be forced to open it up? How are they limiting their competitors?
> they are nearly 2x ahead of Qualcomm using the same ISA at the same power level in 3rd party userspace computation only benchmarks
I'm not sure whether you're speaking about M1 or A14 here. A14 is just a bit faster than other mobile SOCs, certainly not 2x.
If you're speaking about M1... well, what are you comparing it against from Qualcomm? Do they have a similar product, with a similar target and power envelope? The recently announced Snapdragon 8cx Gen 2 from Qualcomm is around 6-7W TDP, which is 50% that of M1.
> There isn't any vertical integration in those, the raw performance is just better.
I'm unconvinced. We need to consider the 5nm production process, the everything-on-soc approach, and the fact that Intel is selling more-or-less the same CPUs as 5 years ago. The Ryzen 7 4800U is quite close to the M1 both performance wise and TDP wise. Now, take a 4800U, make it 5nm, build a SoC with it which includes the memory and build a finely tuned OS - aka "vertical integration". Would you really think you could appreciate a performance or thermal difference?