simt

SIMT stands for Single Instruction, Multiple Threads. It is a parallel execution model used by many modern graphics processing units, most notably NVIDIA’s CUDA-enabled devices. Under SIMT, groups of threads—commonly called warps—execute the same instruction across multiple data elements. The hardware schedules warps on a set of execution units, enabling thousands of threads to participate in data-parallel computations.

Unlike traditional SIMD, SIMT allows threads within a warp to follow different control paths. When a branch occurs, some threads may take the true path while others take the false path. The hardware uses masking to disable lanes not on a given path and re-enables them when threads reconverge. This divergence can reduce efficiency, because differing paths cannot be executed identically in parallel and may require serialized handling within a warp.

In practice, a warp often consists of a fixed number of threads (for example, 32 in many NVIDIA GPUs), while other architectures use different sizes such as 64-thread wavefronts. The scheduler interleaves many warps to hide memory latency and keep execution units busy. Performance is strongly influenced by memory access patterns; coalesced global memory accesses, effective use of shared memory, and caching are important for throughput and latency hiding. Occupancy—the number of active warps per streaming multiprocessor—also affects performance, but is balanced against resource limits per multiprocessor.

SIMT is a core concept in GPU programming models like CUDA and underpins scalable data-parallel computation. It is well suited to workloads such as linear algebra, image processing, and simulations, where many threads operate on large data sets concurrently.