juanrga (noemail.delete@this.juanrga.com) on August 12, 2020 4:22 am wrote:
> ⚛ (0xe2.0x9a.0x9b.delete@this.gmail.com) on August 12, 2020 2:01 am wrote:
> > Adrian (a.delete@this.acm.org) on August 11, 2020 11:00 am wrote:
> > >
> > >
> > > https://nuviainc.com/blog
> > >
> > >
> > > Various sites, e.g. Anandtech, are commenting on this, but that is the primary link.
> >
> > It is probable that advances in µOP cache design, in internal µOP encoding and in x86 instruction
> > fusion will enable AMD/Intel to become competitive with most performance-per-watt competitors claiming
> > that their non-x86 CPU is more efficient because of using the ARM ISA or because of using an ISA with
> > a distance from x86 comparable to the distance between the x86 and the ARM instruction sets.
> >
> > -atom
>
> The efficiency of the ARM ISA is not only in the decoder phase. So magic uop caches will not eliminate the x86-tax.

µOP cache is the storage-part (memory-part) of a binary translation engine. The logic writing data to this storage is located in other parts of the CPU.

ARM A77/A78/X1 have a µOP cache as well. (ARM uses a different terminology here than x86. The terminology hasn't been standardized across the industry yet, formally or informally.)

If ARM's slightly different memory model was a reason for ARM's efficiency then the utilization of L1D cache load/store ports in ARM CPUs would have to be non-trivially better than in x86 CPUs.

-atom