Futures and promises made asynchronous computation easier to compose by giving an unfinished result a value-like placeholder that programs could pass around before the computation completed.
Maurice Herlihy and J. Eliot Moss proposed transactional memory in 1993 as an architectural way to group multiple shared-memory operations into an atomic transaction without conventional locks.
OpenCL emerged in 2008 as an open, royalty-free standard for programming CPUs, GPUs and other parallel processors through one cross-platform execution and memory model.
CUDA removed the need to disguise general computation as graphics work, turning programmable GPUs into a mainstream platform for scientific, technical and later AI workloads.
OpenMP standardized directive-based parallel programming across Fortran, C and C++, giving shared-memory applications a portable path from serial loops to multicore execution.