2026 — NOW
Cryogenic GPU Systems
以 77 K 为目标环境，沿着证据、参数合成、GPU tracing、仿真与设计空间探索建立可复现研究链。
研究 / 77 K
从神经渲染加速器，走向低温 GPU 系统。
当前工作研究 77 K 下器件与系统变化如何重塑 GPU 资源配置和 AI workload。早期神经渲染研究是方法论起点，而不是现在的目的地。
当前问题
低温带来的收益，应该分配到哪里？
研究沿着 evidence-first 链条推进：规范化低温测量、合成有边界的参数、刻画代表性 workload，并在可复现实验中探索架构选择。这里不公开任何未发表结果。
研究演化
2024 — 2026
GPU 架构与系统方法
从单个神经渲染加速器扩展到 workload、memory hierarchy、scheduling 与系统建模。
2022 — 2025
神经渲染硬件加速
围绕 ray representation、sampling、data reuse 与 conflict-free encoding 完成从算法分析到硬件设计的研究。
PUBLICATIONS
代表发表
既有论文记录了此前的神经渲染与加速器阶段。
20262026202520242024202420242023
An Energy-Efficient Edge Coprocessor for Neural Rendering With Explicit Data Reuse Strategies
B. Yuan, X. Zhang, Z. Zheng, Y. Zhang, H. Wan, Z. Yuan, J. Chen, Yunxiang He, J. Ding, X. Zhang, C. Rao, W. Su, P. Zhou, J. Yu, X. Lou
SCAR: A Neural Rendering Accelerator with Sparse-Aware Sampling and Conflict-Free Encoding
Yunxiang He, Yongzhi Zhang, Quanyu Chen, Chaofan Li, Qihan Ding, Xin Lou
A Neural Rendering Coprocessor With Optimized Ray Representation and Marching
Z. Yuan, B. Yuan, C. Rao, Y. Zhu, Yunxiang He, P. Zhou, J. Yu, X. Lou
Density Estimation-based Effective Sampling Strategy for Neural Rendering
Yunxiang He, X. Lou
A 0.59μJ/pixel High-throughput Energy-efficient Neural Volume Rendering Accelerator on FPGA
Z. Yuan, B. Yuan, Y. Gu, Y. Zheng, Yunxiang He, X. Wang, C. Rao, P. Zhou, J. Yu, X. Lou
An Efficient Hardware Volume Renderer for Convolutional Neural Radiance Fields
X. Wang, Yunxiang He, X. Zhang, P. Zhou, X. Lou
Ray Reordering for Hardware-Accelerated Neural Volume Rendering
J. Ding, Yunxiang He, B. Yuan, Z. Yuan, P. Zhou, J. Yu, X. Lou
Analysis and Design of Precision-scalable Computation Array for Efficient Neural Radiance Field Rendering
K. Long, C. Rao, Yunxiang He, Z. Yuan, P. Zhou, J. Yu, X. Lou