Alleviating the main memory bottleneck requires the memory capacity, energy, cost, and performance to all scale in an efficient manner across technology generations.
To avoid hitting the memory limit on GPUs, we devise a batching mechanism to load columns of M based on device’s maximum memory, the results then being accumulated to χ0 from all batches.
Numerous efforts have been undertaken over the past two decades to tackle this challenge, which can broadly be categorized into:
Focus was placed on sharing the two primary (memory consuming) objects, the density and Fock matrices, in the SCF loop among the computation units.
Hence, it is important to shift our focus toward better management of LLC in order to make the most out of it.