Zhufeng (Zephyr) Qiu

Hello — I’m Zhufeng Qiu, and you can also call me Zephyr.

I hold an M.S. in Computer Science from Northeastern University (GPA 4.0), an M.S. in Applied Data Science from University of Southern California, and a B.S. in Geographical Information Science from Wuhan University. Before returning to school I spent a year and a half at NSFOCUS building Kafka/Elasticsearch security-data pipelines and Spark batch workflows serving 100+ enterprise clients at roughly 5M records a day, and I have since been a teaching assistant for Algorithms and Parallel Data Processing at Northeastern.

I’m applying to Ph.D. programs in HPC, ML systems, AI infrastructure, and parallel computing. If you’d like to talk about research or a possible collaboration, email is the fastest way to reach me.

Most of my recent work is the same exercise repeated: take one computation, implement it across every execution model I can reach — serial, threaded, distributed, single-GPU, multi-GPU — hold the numerics exactly constant, and find out what the hardware actually charges for each choice. The interesting results are usually the ones that came out the wrong way round.

Two of them recently did. A lossless 3× compression of an NCCL AllReduce payload bought 4% on NVLink and 55% on PCIe — same binary, opposite verdicts, because the regime decides and not the interconnect’s name. And 4-bit quantization of an 8B model turned out to buy capacity, not speed: NF4 decode runs 1.69× slower than fp16 while moving a quarter of the bytes. Both are written up in projects, with the code and raw measurements public.