Statusvortrag im Promotionsver­fahren von Herrn Hamid Farzaneh

Abstract:

In von Neumann systems, due to the separation of compute and memory, limited bandwidth and high latency create a memory wall where data movement dominates time and energy.  Compute capable memories address this by performing operations close to the data, delivering large gains in speed and efficiency.

One of the reasons for the lag in their adoption is the state of their software stack. Many devices still expose low level device specific controls. Developers must decide when offloading pays off, which memory technology to target, how to place data and map kernels, and which and when optimizations to apply, all under tight device constraints.

Addressing this requires a compilation stack that accepts input from existing high level front ends, introduces intermediate representations to enable optimization and cost models, and applies them to address goals such as target selection, data placement, mapping, and scheduling.