The initial public Mill talks were from mid 2013, when the state of the art was Sandy Bridge, with a 168 instruction reorder buffer and a meek 6-wide execution width, with only 3 ALU capable ports.
Tech has progressed a lot since then. Even with their process issues, Sunny Cove has a ROB twice the size, and the frontend is 11 wide. On the newest node, Apple’s Firestorm has a ~630 ROB, and something like 6 or 7 ALU-capable pipelines, plus another 4 for floating point and vector ops, and all of this using little enough power that a dual-core fits in a smartphone. Things aren’t even slowing down yet; it’s likely in just a few years we’ll be seeing 1k-2k ROBs and correspondingly wide backends.
This makes Silver’s 4 exu slots seem a bit thin. Even the Gold’s 8 exu + 8 flow slots no longer seem particularly wide, especially when the design philosophy is to burn half of them on discarded speculation. Note that I’m not counting reader or writer slots, since they’re mostly implicit in OoO, and I’m not counting picks since they’ll mostly be used for if-conversion. The Mill’s static schedule can’t do lots of things that OoO can, but the response has always been ‘we’re not competing with an OoO of similar width’. Well, now you are, so what’s the plan?
-
This topic was modified 4 years ago by
Veedrac.
-
This topic was modified 4 years ago by
You must be logged in to reply to this topic.