San Francisco Bay Area
MTIA Kernels
Redesigning and implementing the native interface of GraalVM to improve efficiency of native functions calls from GraalVM by leveraging the new Java foreign API.
· Reduced latency by optimizing the mapping of workloads to a custom chip architecture. · Profiled and refactored legacy in-house simulator to bring down execution time from hours to seconds. · Combined heuristics with reinforcement learning for mapping, which decreased latency by up to 13%. · Collaborated with Microsoft engineers to use Bonsai framework for reinforcement learning.