Diploma Thesis

Three implementations of a library used for routing linear algebra sub-problems in multi-GPU systems (with communication - computation overlap).

Implementations

1. CUDA version

2. POSIX threads version

3. HIP version (AMD)

Details

Task queues and an event-based synchronization system are used. The latter implementations extend the applicability of the library and achieve similar or superior performance.

Read my thesis here

© Poutas Sokratis