Jiale Yan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Jiale Yan has a total of 27 co-authors. Here are the top co-authors by number of shared publications:
- Masato Motomura
Institute of Science Tokyo, Yokohama, Kanagawa, Japan
3 shared publications - Anzhou Hou
Li Auto Inc., Shanghai, China
2 shared publications - Daichi Fujiki
Institute of Science Tokyo, Tokyo, Japan
2 shared publications - Kaiwen Long
Li Auto Inc., Shanghai, China
2 shared publications - Kazushi Kawamura
Faculty of Science and Engineering, Waseda University, Tokyo, Japan
2 shared publications - Mo Guang
Li Auto Inc., Shanghai, China
2 shared publications
Latest Publications
- Open Access
- 10 April 2027
- Published by Association for Computing Machinery (ACM)
- p. 874-888
SL²OT: Scaling LLM Locality on Tile with Distributed 3D-DRAM as Unified Memory
Large language models (LLMs) are increasingly limited by memory bandwidth and latency rather than compute throughput. Our modeled 3D-DRAM system offers 4× the raw bandwidth of the HBM reference. However, high remote-access latency (~3× local latency) and redundant NoC traffic reduce its effective bandwidth below that of HBM. We present SL2OT, a hardware--software memory system that harnesses the raw bandwidth of distributed 3D-DRAM to improve LLM performance, guided by four principles---Scalability, Locality, Orchestration, and Temporality. Unified Virtual Memory (UVM) localizes remote accesses through asynchronous prefetching while preserving a unified address space. Asynchronous Request Coalescing (ARC) merges asynchronously arriving requests to the same address across tiles, reducing redundant transfers. Kernel Memory Migration (KMM) prefetches data for upcoming kernels, overlapping host-to-device transfers with computation. Accel-Sim evaluations using A100-80GB SASS traces show 1.22--3.03× speedup over HBM across operators and transformer blocks, and 3.51× speedup over naive distributed 3D-DRAM on GEMM. A matched FlashAttention study reduces mean global-read latency by 73.4% relative to Mesh. SL^2 OT reaches 72.2% effective bandwidth utilization, versus 15.1% without SL2OT.
- 10 April 2027
- Published by Association for Computing Machinery (ACM)
- p. 874-888
Open AccessSL²OT: Scaling LLM Locality on Tile with Distributed 3D-DRAM as Unified Memory
Large language models (LLMs) are increasingly limited by memory bandwidth and latency rather than compute throughput. Our modeled 3D-DRAM system offers 4× the raw bandwidth of the HBM reference. However, high remote-access latency (~3× local latency) and redundant NoC traffic reduce its effective bandwidth below that of HBM. We present SL2OT, a hardware--software memory system that harnesses the raw bandwidth of distributed 3D-DRAM to improve LLM performance, guided by four principles---Scalability, Locality, Orchestration, and Temporality. Unified Virtual Memory (UVM) localizes remote accesses through asynchronous prefetching while preserving a unified address space. Asynchronous Request Coalescing (ARC) merges asynchronously arriving requests to the same address across tiles, reducing redundant transfers. Kernel Memory Migration (KMM) prefetches data for upcoming kernels, overlapping host-to-device transfers with computation. Accel-Sim evaluations using A100-80GB SASS traces show 1.22--3.03× speedup over HBM across operators and transformer blocks, and 3.51× speedup over naive distributed 3D-DRAM on GEMM. A matched FlashAttention study reduces mean global-read latency by 73.4% relative to Mesh. SL^2 OT reaches 72.2% effective bandwidth utilization, versus 15.1% without SL2OT.
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 9343-9352
LearnPrunerV2: Learnable Progressive Token Pruning for Vision-Language Models
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 9343-9352
Open AccessLearnPrunerV2: Learnable Progressive Token Pruning for Vision-Language Models
- Open Access
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1910-1924
BingoGCN: Towards Scalable and Efficient GNN Acceleration with Fine-Grained Partitioning and SLT
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1910-1924
Open AccessBingoGCN: Towards Scalable and Efficient GNN Acceleration with Fine-Grained Partitioning and SLT
- 23 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33 (7) , 1858-1871
DMSA: An Efficient Architecture for Sparse–Sparse Matrix Multiplication Based on Distribute-Merge Product Dataflow
The sparse–sparse matrix multiplication (SpMSpM) is a fundamental operation in various applications. Existing SpMSpM accelerators based on inner product (IP) and outer product (OP) suffer from low computational efficiency and high memory traffic due to inefficient index matching and merging overheads. Gustavson’s product (GP)-based accelerators mitigate some of these challenges but struggle with workload imbalance and irregular memory access patterns, limiting computational parallelism. To overcome these limitations, we propose a distribute-merge product (DMP), a novel SpMSpM dataflow that evenly distributes workloads across multiple computation streams and merges partial results efficiently. We design and implement DMP-based SpMSpM architecture (DMSA), incorporating four key techniques to fully exploit the parallelism of DMP and efficiently handle irregular memory accesses. Implemented on a Xilinx ZCU106 FPGA, DMSA achieves speedups of up to $3.38\times $ and $1.73\times $ over two state-of-the-art FPGA-based SpMSpM accelerators while maintaining comparable hardware resource usage. In addition, compared to CPU and GPU implementations on an NVIDIA Jetson AGX Xavier, DMSA is $4.96\times $ and $1.53\times $ faster while achieving $6.67\times $ and $2.33\times $ better energy efficiency, respectively.
- 23 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33 (7) , 1858-1871
DMSA: An Efficient Architecture for Sparse–Sparse Matrix Multiplication Based on Distribute-Merge Product Dataflow
The sparse–sparse matrix multiplication (SpMSpM) is a fundamental operation in various applications. Existing SpMSpM accelerators based on inner product (IP) and outer product (OP) suffer from low computational efficiency and high memory traffic due to inefficient index matching and merging overheads. Gustavson’s product (GP)-based accelerators mitigate some of these challenges but struggle with workload imbalance and irregular memory access patterns, limiting computational parallelism. To overcome these limitations, we propose a distribute-merge product (DMP), a novel SpMSpM dataflow that evenly distributes workloads across multiple computation streams and merges partial results efficiently. We design and implement DMP-based SpMSpM architecture (DMSA), incorporating four key techniques to fully exploit the parallelism of DMP and efficiently handle irregular memory accesses. Implemented on a Xilinx ZCU106 FPGA, DMSA achieves speedups of up to $3.38\times $ and $1.73\times $ over two state-of-the-art FPGA-based SpMSpM accelerators while maintaining comparable hardware resource usage. In addition, compared to CPU and GPU implementations on an NVIDIA Jetson AGX Xavier, DMSA is $4.96\times $ and $1.53\times $ faster while achieving $6.67\times $ and $2.33\times $ better energy efficiency, respectively.
- Open Access
- 30 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 10398-10411
TT-MLP: Tensor Train Decomposition on Deep MLPs
Deep multilayer perceptrons (MLPs) have achieved promising performance on computer vision tasks. Deep MLPs consist solely of fully-connected layers as the conventional MLPs do but adopt more sophisticated network architectures based on mixer layers composed of token-mixing and channel-mixing components. These architectures enable deep MLPs to have global receptive fields, but the significant increase of parameters becomes a massive burden on practical applications. To tackle this problem, we focus on using tensor-train decomposition (TTD) for compressing deep MLPs. At first, this paper analyzes deep MLPs under conventional TTD methods, especially using various designs of a macro framework and micro blocks: The former is how to concatenate mixer layers, and the latter is how to design a mixer layer. Based on the analysis, we propose a novel TTD method named Train-TTD-Train . The proposed method exerts the learning capability of channel-mixing components and improves the trade-off between accuracy and size. In the evaluation, the proposed method showed a better trade-off than conventional TTD methods on ImageNet-1K and achieved a 0.56% higher inference accuracy with a 15.44% memory reduction on Cifar-10.
- 30 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 10398-10411
Open AccessTT-MLP: Tensor Train Decomposition on Deep MLPs
Deep multilayer perceptrons (MLPs) have achieved promising performance on computer vision tasks. Deep MLPs consist solely of fully-connected layers as the conventional MLPs do but adopt more sophisticated network architectures based on mixer layers composed of token-mixing and channel-mixing components. These architectures enable deep MLPs to have global receptive fields, but the significant increase of parameters becomes a massive burden on practical applications. To tackle this problem, we focus on using tensor-train decomposition (TTD) for compressing deep MLPs. At first, this paper analyzes deep MLPs under conventional TTD methods, especially using various designs of a macro framework and micro blocks: The former is how to concatenate mixer layers, and the latter is how to design a mixer layer. Based on the analysis, we propose a novel TTD method named Train-TTD-Train . The proposed method exerts the learning capability of channel-mixing components and improves the trade-off between accuracy and size. In the evaluation, the proposed method showed a better trade-off than conventional TTD methods on ImageNet-1K and achieved a 0.56% higher inference accuracy with a 15.44% memory reduction on Cifar-10.