SIMD

SIMD stands for single instruction, multiple data, a computing paradigm in which the same operation is applied simultaneously to multiple data elements. By exploiting data-level parallelism, SIMD can increase throughput for workloads with repetitive operations on large data sets, such as multimedia processing, graphics, and numerical simulations.

Most modern CPUs include vector processing units that operate on wide registers (for example 128-, 256-, or 512-bit). A single vector instruction can perform the same operation on multiple lanes, such as adding two vectors element-wise or comparing elements in parallel. Data alignment and memory bandwidth are important considerations, as misaligned data or cache misses can negate the benefits.

Historically, SIMD originated in vector supercomputers and later became mainstream on general-purpose processors. Intel's SSE and AVX, ARM's NEON, and PowerISA's AltiVec/VMX are widely used SIMD extensions. Some architectures also support gather/scatter operations to handle non-contiguous data. GPUs employ a related, but distinct, approach called SIMT, which groups many threads executing the same instruction.

Software support includes compiler auto-vectorization, intrinsics for explicit control of vector instructions, and high-level libraries or languages that expose SIMD capabilities. Portable vector APIs and ongoing standardization efforts aim to improve cross-platform code reuse.

Limitations of SIMD include the need for data parallelism, overhead from data packing/unpacking and type conversions, and diminishing returns when branches or irregular memory access patterns dominate. Effective use often requires data layout changes, loop tiling, or algorithm redesign.