They can use whatever they want, it does not mean that the code optimally uses the hardware. For example, if you code uses a naive loop with vector instructions to accumulate a long chain of numbers, an x86 CPU with AVX2 will win - it can do twice as many additions per cycle and it runs higher clocks, and your execution is limited by the dependency chains. But if you unroll your code to allow multiple instructions to execute simultaneously, M1 will get ahead as it has more vector processing units and better memory subsystem.
This is not true. Benchmarks show M1 outperforming anything at the same power consumption.
Complete nonsense. I might agree that for chess engines, the performance is slightly disappointing. For most other applications, M1 has demonstrated high performance. Industry-standard benchmarks, real-world use as well as specialized tests make it very clear. You can’t cherry pick a single domain and ignore everything else.
AMD 5700G uses 5x as much power than M1 and has a slower GPU. It is slower than M1 in single core and 30% faster in multi-core, despite having twice as many cores and using much more power. Yes, performance per dollar is very good. If you want performance per dollar, there are much better options out there than Macs.
M1 does not need AVX2 since it has double the number of vector compute units than x86 CPUs. Modern x86 can do 2 256-bit AVX2 operations per cycle, M1 can do 4 128-bit Neon operations per cycle. And M1 sind operations have lower latency. There are plenty of tests showing that M1 can hold its own agains AVX512.
M1 does not need VNNI since it offers multiple extensions for machine learning. It has the AMX coprocessor that is capable of performing 256 FP32 operations per cycle. A modern Intel AVX512 core can only do 16 operations per cycle. In other words, a single M1 has matrix multiplication throughput comparable of that of 16 Tiger Lake cores. Not to mention that M1 has an additional matrix multiplication unit (the NPU), which is still faster.
Who cares? Hyperthreading is a hack designed to squeeze a bit more multithreaded performance of a CPU that is unable to efficiently utilize its execution resources. M1 can have three times more instructions in flight than fastest x86 cores, it won’t benefit from hyperthreading.
Write better code. If your program runs slower on M1, you are probably doing something very wrong.
Last edited:
