Large-scale AI Model distributed training- Analyzed bottlenecks using a profiler in AI models for molecular dynamics simulations such as DimeNet, GemNet, NequIP, etc., and explored optimization strategies- Utilized a distributed training library with up to 1024 GPUs to train a large language model and found the most optimal configuration- Contributed to advancing the software culture as a certified SW Professional, Associate Architect, and Best ReviewerProcessing in Memory- Designed and implemented a PIM software stack enabling users to execute unmodified source code built on popular machine learning frameworks including Tensorflow and PyTorch- Developed GPU kernel code to activate the PIM feature in High Bandwidth Memory (HBM) embedded on the GPU- Successfully enabled system and computation using PIM, a world-first achievement, with real GPU hardware- Accelerated performance approximately 1.5x to 2.0x for natural language processing models such as DeepSpeech2, RNN-T, and GNMT during inference