Yang Zhou

Yang Zhou

Auto-generated profile, based on publicly available information

Yang Zhou

Tsinghua University, Beijing, China

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
11 February 2025 in IEEE Transactions on Circuits and Systems II: Express Briefs
Verified
Journal Article
Closed access
YZYang Zhou
Yang Zhou
YWYang Wang
Yang Wang
WYWenxin Yin
Wenxin Yin
YJYazheng Jiang
Yazheng Jiang
JWJingchuan Wei
Jingchuan Wei
SYShouyi Yin
Shouyi Yin
Previous digital Compute-In-Memory (DCIM) exhibits limitations in Deep Neural Network (DNN) training with FP8, which is playing an increasingly important role in model training. Most of the previous DCIM purely align exponents to the maximum one when computing floating-point data. This method struggles to balance accuracy and energy efficiency when using FP8. In this paper, we propose BeaCIM, a DCIM processor with bi-directional exponent alignment. There are several contributions. First, we propose a new exponent alignment mechanism, which can dynamically adjust the shared exponent (E*) towards the numerical distribution center of exponents. Second, we use the Shift-on-Product (SOP) method to address the limitation of data bitwidth, and present E* calculator based on the Ordinary Least Squares (OLS). Finally, we performed RTL-level circuit implementation and evaluated BeaCIM using various datasets and models. Our experiments show the following results: (1) Compared with the standard hybrid FP8 training, our bi-directional exponent alignment FP8 training exhibits an average top-1 accuracy drop less than 0.75% across different models and datasets. (2) BeaCIM can achieve an energy efficiency of 31.5 TFLOPS/W at FP8, which is 2.6-19.3× better than state-of-the-art works.

1 January 2026 in Electronics Letters
Verified
Journal Article
Open access
HHHong Hu
Hong Hu
ZLZhiwei Lin
Zhiwei Lin
YZYang Zhou
Yang Zhou
XYXiaolong Yang
Xiaolong Yang
WXWende Xu
Wende Xu
SYShouyi Yin
Shouyi Yin
Floating‐point (FP) compute‐in‐memory (CIM) is promising for accelerating deep neural networks (DNNs). To support emerging microscaling (MX) FP formats without severe accuracy degradation, FP‐CIMs must adopt a post‐alignment dataflow. However, conventional post‐alignment dynamically detects the maximum exponent () with a power‐hungry multi‐stage comparator tree, introducing substantial comparison energy and area overheads. This letter proposes a quasi‐exponent shift alignment (QESA) mechanism for energy‐efficient post‐aligned FP‐CIMs. QESA targets a quasi‐maximum exponent () rather than the exact , and extracts it using a lightweight bit‐wise self‐gating OR circuit. A compact bi‐directional shifter then performs ‐guided alignment, followed by post‐normalization, achieving outputs mathematically equivalent to those of exact ‐based alignment. Implemented in a 28 nm CMOS process, QESA reduces the area and power of the detector by 86.3% and 89.7%, respectively, at the component level; at the subsystem level, it decreases the adder tree energy by an average of 28.2% across the evaluated workloads.

28 April 2026 in IEEE Open Journal of the Solid-State Circuits Society
Verified
Journal Article
Open access
YZYang Zhou
Yang Zhou
ZLZhiwei Lin
Zhiwei Lin
YQYubin Qin
Yubin Qin
YHYang Hu
Yang Hu
YWYang Wang
Yang Wang
SYShouyi Yin
Shouyi Yin
Large language models (LLMs) have achieved remarkable success across vision, text, and video domains, far surpassing traditional neural networks. However, this success is accompanied by enormous memory access and computational demands. Quantization is a widely used technique to reduce model overhead. Microscaling is a novel fine-grained quantization scheme that employs finer block-wise partitioning, delivering superior performance compared to traditional methods like per-tensor quantization. However, the widespread presence of outliers in models still poses challenges for microscaling quantization. First, a fixed block-partitioning pattern with a single dimension cannot adapt to the random distribution of outliers, resulting in degraded quantization performance. Second, outliers within a block cause excessive numerical error in low-bit quantization, forcing the use of higher bit-widths and resulting in wasted power. To address these challenges, first we propose a bidirectionally partitioned microscaling method that dynamically selects the partitioning direction with fewer outliers. Second, we separate outliers and non-outliers into sparse and dense data, allowing most non-outlier data to be quantized to low bit-widths effectively. Third, we designed dense/sparse hybrid cores to separately process non-outlier and outlier data. Experimental results show that our approach reduces the outlier count by 35.9% and memory access by 66.7% on average. The proposed accelerator achieves a peak energy efficiency of 9.81 TFLOPS/W at MXFP4, and outperforms state-of-the-art accelerators with 1.49× lower energy and 1.38× higher speed.