Scholar
Mieszko Lis
Google Scholar ID: HaniDJ8AAAAJ
The University of British Columbia
Citations & Impact
Citations
734
H-index
14
i10-index
23
Publications
20
Co-authors
0
Contact
Publications
Resume
Academic Achievements
- Developed novel side-channel attacks on sparse DNN accelerators capable of recovering model architectures despite encryption
- Designed an algorithm and hardware architecture reducing DNN training time and energy by 4× at iso-accuracy
- Proposed convolutional layer fusion technique achieving up to 6× performance improvement
- Integrated resistive memories with DNN accelerators for up to 6× energy savings
- Created scheduling algorithms for tensor workloads that generate energy-efficient schedules faster than prior tools
- Implemented efficient strong memory semantics and hardware transactional memory in GPUs
- Used locality-sensitive hashing to significantly improve cache compression ratios and advocated decoupling compression from approximation
- ML-based prefetcher placed 2nd in the 2021 ML-Based Data Prefetching Competition
- Developed an abstraction and toolchain to factor data structures and reduce cache misses
- Designed an FPGA MAP equalizer supporting Faster-than-Nyquist signaling for 5G
- Built fast, accurate static/dynamic program slicers for bug tracking in Android and Java
- Invented a dynamic, path-aware taint analysis technique
- Realized the 110-core EM² chip validating execution migration for shared memory
- SSC enabled the nanomedicine game NanoDoc
Research Experience
- Worked in industry prior to UBC on Bluespec and high-speed network processors
- Led research on DNN accelerator attacks/defenses, efficient DNN training/inference architectures, GPU memory semantics, cache compression/prefetching, split-frame rendering, 5G FPGA equalizers, program slicing, and taint analysis
- Developed the Execution Migration Machine (EM²), a 110-core 45nm ASIC implementing unified shared memory via execution context migration
- Created the scalable NoC simulator HORNET and researched oblivious routing, multi-path in-order delivery, and VC scheduling
- Designed SSC, a domain-specific language and compiler for biochemical reaction networks, applied to immunology and cell-membrane signaling
- Built Spanner, a protein structure modeling tool for threading sequences onto 3D templates