演讲人：Jung Ho Ahn [首尔国立大学]
时间：10:00-11:00, Nov 7, 2025 (Fri)
地点：RM 1-222, FIT Building
内容：
GPUs continue to redefine performance boundaries across diverse workloads. This talk presents two compelling examples from our latest research, underscoring the critical role of programmable accelerators. First, we demonstrate how our group surpassed DARPA's ambitious 2021 DPRIVE ASIC target for fully homomorphic encryption (FHE). Using an NVIDIA RTX 5090, we achieved 22.4ms (cf., the ASIC target of 25ms) latency with 87% accuracy for private 7-layer CNN inference on CIFAR-10---without ASIC fabrication---through our innovative software and algorithmic co-design (leveraging insights from projects like AESPA and HyPHEN). Second, we critically re-evaluate the evolving landscape of Processing-In-Memory (PIM) for LLM inference. While our ASPLOS 2024 work (AttAcc) highlighted HBM-PIM's potential for low arithmetic intensity MHA (~1), we now show how newer attention variants (GQA ~10, MLA ~100) shift the paradigm, (undermining the need for HBM-PIM, showcasing the potential for custom HBM, and) further emphasizing the need for highly programmable hardware in the fast-paced domains of FHE and Generative AI.
个人简介:
Jung Ho Ahn is a professor at Seoul National University. Professor Ahn received his Ph.D. in electrical engineering from Stanford University (2007), was a senior research scientist at HP labs before joining Seoul National University, and took a sabbatical at Google (2016) and Samsung (2024). His research interests include bridging the gap between the performance/efficiency demand of emerging applications and the potential of modern and future massively parallel systems, more specifically on memory sub-systems. Professor Ahn is the hall of fame member of HPCA, ISCA, and MICRO.
