Konrad Schwarz (no.spam.delete@this.no.spam) on August 30, 2018 6:37 am wrote:
> Maynard Handley (name99.delete@this.name99.org) on August 18, 2018 4:22 pm wrote:
> > - you offer an atomic remote zero (or remote fill) instruction that is given a line address
> > plus count. The first thing it does is flush the line from all other caches, then it marks
> > the set of lines locally as in some sort of intermediate state, then, fast as possible, it
> > starts zeroing each line and, once zeroed, it goes back to a standard "valid" state.
>
> Power(PC) Data Cache Block Allocate (dcba) and Data Cache Block
> Zero (dcbz) instructions do this, albeit on a single cache line.

I'm well aware of those (and used them extensively back when I was at Apple).
But they're not perfect for what I had in mind.
They are good for reducing the overhead of writeback, but they still allocate in the local L1.
So if you have some sort of streaming scenario (eg zeroing anything longer than 32K or 64K bytes) or even you know that the page you've zero'd won't be used soon, then you're still pointlessly paying the cost of moving the data back out from the L1.

Shipping (a small set of specific) instructions to data rather than vice versa has, of course, been proposed one way or another for many years. But I suspect that the companies with their eyes on the ball will start to take it more seriously.
If the questions are: "how do I go faster, in the face of static transistor performance at smaller nodes" OR "what sort of extra value can I create to justify going to smaller nodes (ie using the extra transistors available)" these sorts of lateral moves out of the core proper make more and more sense.

A variant of this is the idea of certain compute primitives available in the SSD. I know Samsung released this sort of thing a few years ago, and I've seen some (not-entirely academic) discussions of how well it can work for certain purposes, but the targeted applications were, naturally, enterprise and not, per se, of much interest to me; so I've no feel what sort of a more generic future this particular version of the idea has.