Vendor-specific parallel computing.
WHAT IT IS
A platform for general computation on one vendor's graphics processors.
WHY IT DOMINATES
Mature tooling, extensive libraries, and the ecosystem built around it.
WHAT THE MODEL IS
Kernels executed by many threads, organised into blocks and grids.
WHAT A BLOCK PROVIDES
Threads that can share fast memory and synchronise with each other.
WHAT THREADS IN DIFFERENT BLOCKS CANNOT DO
Synchronise, within a kernel.
WHAT TO CHOOSE CAREFULLY
Block size, which affects how well the hardware is occupied.
WHAT THE LIBRARIES PROVIDE
Linear algebra Fourier transforms Deep learning primitives Sorting and scanning
WHAT TO USE FIRST
Those libraries.
WHY
They are heavily optimised, and hand-written kernels rarely match them.
WHAT TO PROFILE WITH
The vendor's profiling tools, reporting occupancy, memory throughput and divergence.
WHAT COMMONLY LIMITS PERFORMANCE
Memory access patterns Low occupancy Transfers dominating
WHAT TO MEASURE
End-to-end time, including transfers.
WHAT THE LOCK-IN IS
Code written for it runs only on that vendor's hardware.
