[jimmyed]

> When one sibling is executing, the other must wait. SMT is useful because CPUs often go idle while waiting for events — usually the arrival of data from memory. While one CPU waits, the other can be executing

Is this accurate? I was under the impression SMT gains are not from running other threads when one is blocked (preemption is a old feature) but the processor having a multi stage pipeline so that the net number of instructions that are executed per cycle is more than 1 (closer to 2 in the above example)

[Symmetry]

No, it's not. SMT is specifically when you're running two threads simultaneously on a superscalar processor that can execute multiple instructions in a single cycle. During SMT one thread might be using execution reasources the other thread doesn't need further increasing performance beyond just having something to do during a cache miss.

What the article describes is temporal multithreading where only one thread executes at a time.

https://en.wikipedia.org/wiki/Multithreading_(computer_archi...

[willis936]

As someone else said, this is superscalar, though these days companies refer to the number of concurrent out-of-order pipelines as Execution Units (EUs).

SMT is a second set of registers that uses the same execution hardware. If a thread stalls waiting for a main memory read then the other thread can jump in after the pipeline is cleared out (or maybe before if you're clever and careful, but idk if this is done in practice). You can also impose some governance over time sharing so threads with low data memory footprints don't lock out other threads.

[zaxcellent]

What you're describing sounds like a superscalar processor which utilizes instruction level parallelization: https://en.wikipedia.org/wiki/Superscalar_processor