Systems Dictionary • Hardware Co-Design • SIMD
SIMD Vectorization
A parallel hardware execution model where a single processor instruction simultaneously performs the same mathematical or bitwise operation across multiple data points packed into wide vector registers.
Hardware Architecture Mechanics
Standard scalar CPU execution loads one 32-bit or 64-bit value into a general-purpose register and computes one result per cycle. SIMD introduces wide registers (128-bit SSE/NEON, 256-bit AVX2, or 512-bit AVX-512) that hold multiple elements:
Scalar: [ a0 ] + [ b0 ] = [ c0 ] (1 op per instruction)
AVX2: [ a0 | a1 | a2 | a3 | a4 | a5 | a6 | a7 ] + (8 x 32-bit floats)
[ b0 | b1 | b2 | b3 | b4 | b5 | b6 | b7 ] =
[ c0 | c1 | c2 | c3 | c4 | c5 | c6 | c7 ] (8 ops in 1 clock cycle)
AVX2: [ a0 | a1 | a2 | a3 | a4 | a5 | a6 | a7 ] + (8 x 32-bit floats)
[ b0 | b1 | b2 | b3 | b4 | b5 | b6 | b7 ] =
[ c0 | c1 | c2 | c3 | c4 | c5 | c6 | c7 ] (8 ops in 1 clock cycle)
In database engineering and high-speed JSON parsing (such as simdjson), SIMD scans 64 characters of text in a single CPU cycle to identify structural delimiters (braces, quotes, colons) without branching.
Modern Vector Instruction Sets
| Architecture | Register Width | Throughput (Bytes/Cycle) | Primary Target |
|---|---|---|---|
| ARM NEON | 128-bit | 16 bytes | Apple Silicon (M1-M4), AWS Graviton |
| Intel AVX2 | 256-bit | 32 bytes | Modern x86-64 Servers (AMD EPYC, Intel Xeon) |
| Intel AVX-512 | 512-bit | 64 bytes | Enterprise Xeon, AMD Zen 4/5 Compute Nodes |