SPMD

SPMD is a programming model, not hardware level organisation (like SIMD or MIMD)
See Programming Model

NOTE: A “for loop; running over 1M iterations; performing X operatios in each iteration” is not an example of SPMD programming, but Sequential Programming.
”A multi-threaded application; ran over 1M threads; each thread performing X operations” is an example of SPMD programming.

The SIMT programming model of GPUs is basically SPMD implemented on SIMD hardware (array and vector processors).

Even before GPUs or SIMD processors (1990s), people used to use SPMD programming model to run the same program across multi-processors (MPI applications)
But running same program across different processors has an overhead of instruction cycle (fetch, decode, execute).

Since underlying hardware of a GPU is a SIMD, GPUs are much more efficient, even when using the same programming model.

SPMD vs SIMD programming models

SPMD differs from SIMD programming in the same way a program and an instruction are different.

Program is a group of smaller commands, where each command is an Instruction.
Instructions are tied to the underlying architecture. But a program is not.

Also, SPMD does not guarantee that same instructions will be performed on all the data elements, since different execution path is possible due to control flow.
SIMD guarantees that same instruction is performed on all data elements