See the Data Streaming Accelerator:
https://01.org/blogs/2019/introducing-intel-data-streaming-a...
Ah, a new name for IOAT (PCIe enumerated DMA engine in the SoC). I used this back on Broadwell.
Re:Fast memcpy, A System Design
by userbinator on Monday December 19, 2022 @07:03PM
(#34059676)
what is Intel (or AMD) doing regarding this?
In this case they haven't invented new instructions at all but have been steadily improving REP MOVS over time.
So does that mean that further major optimizations are basically impossible?
You don't always need the operation finished right away. Maybe REP MOVS could return a sort of a handle that you then wait for when you actually need to use the destination, and the CPU can keep chugging along in the background. Like when you submit commands to the GPU or how you can wait at memory fences for other threads in a warp to complete.
Further optimization are certainly possible, especially as the hardware improves and gains additional capabilities. With regards to your suggestion, this is essentially what already happens, thanks to the magic of out-of-order cores.
Surely they are aware of the memcpy problem, so if it's true that servers spend a lot of time copying memory, what is Intel (or AMD) doing regarding this?