Taiga Kobayashi
Advanced Computing ACceleration (AC2) Laboratory (PI: Assoc. Prof. Ryohei Kobayashi)
I work on communication optimization for large-scale systems, with interests in FPGAs, DPUs, machine learning, and data compression. My current work studies communication-data compression on NVIDIA BlueField DPUs for multi-GPU LLM training, with an eye toward communication substrates for future sparse and irregular accelerator workloads.
01Research
Characterizing communication compression on DPUs for distributed training
Implemented a NetZIP-style lossless pipeline (byte grouping + LZ4, 16-core parallelism, chunk-level compression/transfer overlap) on the Arm cores of an NVIDIA BlueField-3 DPU, and compared it end-to-end against uncompressed transfer on the same RDMA path. Quantified that the uncompressed baseline shifts with RDMA-buffer page backing (4 KiB: 10.9 GB/s vs. THP: 24.6 GB/s), and that sender-side processing saturates at about 9 GB/s — yielding concrete design implications.
DPURDMAlossless compressiondistributed trainingResearch interests
FPGA / DPU / Machine Learning / Data Compression
02Publications
Taiga Kobayashi, Tomohiro Ueno, Ryohei Kobayashi, “Characterizing Communication Compression on NVIDIA BlueField-3 DPU for Distributed Training,” Workshop Proceedings of the 55th International Conference on Parallel Processing (ICPP Workshops ’26), SPIN-NVSC Workshop, Singapore, Sep. 2026 (to appear). DOI
Taiga Kobayashi, Tomohiro Ueno, Ryohei Kobayashi, “Characterization of communication compression on NVIDIA BlueField-3 DPU for distributed training” (in Japanese), IPSJ SIG Technical Report (SWoPP 2026, SIG-HPC), Jul. 2026. IPSJ DL
03Education
-
Apr. 2022Entered the Bachelor's program, Institute of Science Tokyo
-
Mar. 2026Bachelor's degree, Institute of Science Tokyo
-
Apr. 2026 – presentMaster's program, Department of Mathematical and Computing Science, School of Computing, Institute of Science Tokyo (expected Mar. 2028); Research Assistant, JST ASPIRE project "Exploration of Adaptive Data-Compression Hardware for Next-Generation High-Efficiency Computing Infrastructure" (PI: Tomohiro Ueno, RIKEN Center for Computational Science)