Award-Winning HPCA 2025 Papers on Near‑DRAM Processing (UniNDP) and GPU‑Accelerated Fully Homomorphic Encryption (WarpDrive)

In the era driven by artificial intelligence and big data, efficient processing of massive computational tasks places higher demands on hardware architectures, with near‑storage computing and fully homomorphic encryption (FHE) emerging as frontier technologies to address memory‑wall bottlenecks and privacy‑preserving computation challenges.

The International Symposium on High‑Performance Computer Architecture (HPCA 2025) was held in Las Vegas, USA, from March 1‑5, 2025, accepting only 113 papers (21.2% acceptance rate). Two research works from Tsinghua University and Ant Group’s research institute were highlighted.

Paper 1: UniNDP – A Unified Compilation and Simulation Tool for Near‑DRAM Processing Architectures proposes a unified compilation‑simulation toolchain based on abstracted architecture and instruction set models, enabling efficient deployment of neural‑network workloads across diverse near‑storage architectures. Compared with existing data‑mapping and compilation methods, UniNDP achieves up to 3.43× end‑to‑end speedup and received a Best Paper Honorable Mention.

Paper 2: WarpDrive – GPU‑Based Fully Homomorphic Encryption Acceleration Leveraging Tensor and CUDA Cores introduces a GPU‑accelerated FHE solution that exploits Tensor Core‑based Number Theoretic Transform (NTT) implementations and fine‑grained warp‑level memory access designs, reducing instruction count and pipeline stalls. WarpDrive delivers up to 13.3× NTT throughput improvement and overall speedups of 13.4×, 3.5×, and 2.8× for NTT, homomorphic operations, and applications respectively.

The session also includes a live “Paper Showcase” where the first authors, Xie Tongxin (Tsinghua University) and Fan Guang (Ant Group), will share design ideas and verification processes. The live broadcast is scheduled for March 20, 2025, 18:00‑20:00 on WeChat Video, AntTech, SecretFlow, and Bilibili platforms.