Comment by zeusk

Comment by zeusk 4 days ago

1 reply

SMT isn't that really is it?

SMT allows for concurrent execution of both threads (thus independent front-end for fetch, decode especially) and certain core resources are statically partitioned unlike a warp being scheduled on SM.

I'm not a graphics expert but warps seem closer to run-time/dynamic VLIW than SMT.

petermcneeley 3 days ago

In actual implementation they are very much like very wide SIMD on a CPU core. Each HW thread is a different warp as each warp can execute different instructions.

This mapping is so close that translation from GPU to CPU relatively easy and performant.