So, I have done some digging, into that Amazon's lately announced Graviton2 CPU, that the debate has been going on for few last pages.
It is based on ARM Neoverse Cores.
https://www.arm.com/products/silicon-ip-cpu/neoverse/neoverse-e1
And this is the point where the whole theory of ARM being faster than x86 falls apart.
Neoverse cores are 15-20% faster per core, than last gen. They are still ARMv8 architecture, the same as Graviton 1 cores. They are designed for higher scalability however, up to 128 cores in one chip, but they lack Hyper Threading/SMT.
So IPC of Graviton 2 CPUs rose at best 20-25%, with any custom work done on the CPU. Most of the benefits most likely come from dedicated hardware on chip, which Amazon, most likely included into the transistor budget - 30 bln xTors on 7 nm and only 32 cores? Thats like 2x more transistors compared to 32 Core 7 nm Threadripper 3970X. Where did they went? The pipeline is not long enough, considering the clocks of the CPU, so Amazon hasn't burned it on increasing the clocks, most likely what did they burned them for is dedicated ASIC on chip, for specific tasks liek Machine Learning or Image Processing, which is why they claim it is 20% faster than Unspecified Intel CPU in an unspecified testing methodology. What if they tested singlecore performance of 180W TDP CPU that is locked to 105W TDP, and comkpared it to a CPu that was designed to operate at that TDP. How clock speeds would behave on those CPUs, hm?
And for last thing: I suggest reading this post on Phoronix:
Phoronix: Amazon Talks Up Big Performance Gains For Their 7nm Graviton2 CPUs We weren't too enthusiastic about the performance of Amazon's initial Graviton ARM-based CPU cores offered via their Elastic Compute Cloud, but their next-gen Gravin2 CPUs that are "coming soon" should be much more...
www.phoronix.com
I hope this end this ridiculous debate about ARM vs x86. There is nothing more powerful in high-performance than x86, and for the foreseeable future there won't be. Period.
wccftech.com
