> > > The essential insight is that
> > > - most memory requests are strongly clustered to a single page or a few pages
> > > - to the extent that requests are spread over many pages, that same code
> > > is probably not exactly latency essential to the single cycle level
> >
> > doesn't sound like modern real world workloads to me, tbh
>
> The link
>
> https://eprints.soton.ac.uk/347147/1/__userfiles.soton.ac.uk_Users_spd_mydesktop_MALEC.pdf
>
> works for me.
>
> Here's an alternative version, or you can search for the title:
> MALEC: A Multiple Access Low Energy Cache
>
> (The paper gives *some ideas* that are used by the M1. It is not a description
> of the M1's cache, which also uses other very different ideas.)
>
> As for your claims about "modern workloads", well, what can I say?
> The M1 works this way. And it works very well for Apple's workloads.
> You want to deny these two facts, go ahead; I care about explaining, not converting.

I simply disagree with the two bullet items -- single/few pages, not latency essential.

Because that's not what I see when I study modern real life workloads -- they are big, and they love pointer chasing.

And seeing which benchmarks the paper authors studied, their conclusion doesn't surprise me.

That is all.