Abstract

Increasing hardware bandwidth (PCIe, high-speed storage, and high-bandwidth memory) has shifted performance bottlenecks from I/O to the CPU, making aggressive optimization essential even for seemingly simple tasks. The simdjson C++ library that leverages SIMD instructions to parse JSON at speeds up to 14 GB/s on a single core of a macBook, using a fast indexing technique to identify structural elements and enabling efficient operations such as validation, skipping, and minification. The library now incorporates C++26 compile-time reflection for zero-overhead, boilerplate-free serialization and deserialization of user-defined types. We also present the simdutf library, which delivers dramatic speedups for Unicode transcoding (UTF-8/UTF-16 validation and correction) and Base64 encoding/decoding—reaching over 16 GB/s in browsers—capabilities now integrated into V8 and WebKit. We review the evolution of SIMD capabilities across CPU architectures (x86, ARM, RISC-V, etc.), the importance of portable yet hardware-aware optimizations in C++, and the value of open-source community collaboration in pushing performance boundaries.