Thrilled to announce that our research group (scale@snu) kicks off 2025 with a paper accepted at HPCA (besides two conditionally accepted papers @ ASPLOS 2025). All streaks will end; but we are still enjoying a streak, 11 in a row, of publishing at least one paper to ISCA/MICRO/HPCA since MICRO 2021. MICRO#4 (TRiM) - HPCA#10 (Mithril) - ISCA#9 (BTS) - MICRO#5 (ARK) - HPCA#11 (SHADOW) - ISCA#10 (SHARP) - MICRO#6/#7 (Cube/CXL-DRAM) - HPCA#12 (CXL-PNM) - ISCA#11/#12 (DRAMScope/ND$) - MICRO#8 (Duplex) - HPCA#13 (Anaheim).

In this Anaheim paper (“Anaheim: Architecture and Algorithms for Processing Fully Homomorphic Encryption in Memory”), we investigated the role of NDP/PIM (near data processing, processing in memory) in accelerating FHE (fully homomorphic encryption, the CKKS scheme in particular). For this purpose, we first rewrote a GPU library for the CKKS FHE, called Cheddar (it is substantially better than what we have developed and published through the 100x paper; it is arxived as of now, but we hope that it will be peer-reviewed soon as well).

Through this refactoring process, we identified that an optimized GPU implementation leaves quite a bit of room for further enhancement by exploiting NDP/PIM. This is because over half of the execution time is bottlenecked by memory-intensive kernels, mostly performing element-wise operations. We believe that FHE acceleration can take advantages of NDP/PIM, which is well aligned with the trendy buzzwords, such as Custom HBM and LPDDR-PIM (On a related note, I am taking a sabbatical at Samsung DS now; I am more than glad to have a coffee or lunch with many of the Samsung employees who know me directly/indirectly!).

Anaheim follows the trilogy papers from my group regarding CKKS acceleration -- BTS/ARK/SHARP (CiFHER might have shared its position constituting a quad, but we will work on further advancements regarding chiplet for FHE acceleration). However, Anaheim takes a bit different stance as it leverages existing GPUs for compute-intensive operations.

I appreciate the heroic efforts led by Jongmin Kim, the first author of Anaheim (and ARK/SHARP), and the team (Sungmin Yun, Hyesung Ji, Wonseok Choi, and Sangpyo Kim). He and the gang turned an ambiguous idea and an ambitious goal set by me into a gem. The critical and constructive comments given by the reviewers and the extensive revision efforts made by the team turned the lukewarm review scores into all positive ones. I truly hope that the efforts made by us and the other computer architecture community help homomorphic encryption approaching a broader acceptance.