SIMT: คำสั่งเดียว หลายเธรด

Thousands of cores need a smart way to be told what to do. The GPU's answer is SIMT: Single Instruction, Multiple Threads. ⚡

In SIMT, one instruction is broadcast to a whole group of threads at once. Each thread runs that same step, but on its own piece of data.

Picture a kitchen where every cook follows the exact same recipe step, but each works on a different plate. That shared step is your instruction. 🍳

The GPU groups threads into bundles of 32 called a warp. A warp is the real unit that executes together, lockstep, one instruction at a time.

Issuing one instruction for 32 threads at once is far cheaper than 32 separate commands. That sharing is exactly where the GPU's efficiency comes from.

Threads in a warp share the instruction but keep private registers. So thread 0 and thread 5 run the same add, just on different numbers.

Classic SIMD processes fixed-width vectors. SIMT keeps the idea of shared instructions but lets each thread behave more independently when needed.

What if half a warp takes an if branch and half does not? Threads in a warp want to march together, so a branch can split the group apart.

When threads in a warp disagree on a branch, the warp runs each path in turn and disables the others. This serial replay is called divergence.

Because divergent paths run one after another, you lose parallelism. Keeping a warp on the same path is a key idea for fast kernels.

With one instruction feeding 32 threads, and many warps in flight, the GPU keeps its math units packed. That is how SIMT turns into raw throughput.

Let us make sure the SIMT vocabulary is solid.

SIMT broadcasts one instruction to a warp of 32 threads, each on its own data. Avoid divergent branches to keep every thread marching together. 👍