arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2603.08712cs.AR
收藏

一种具有形式误差界限的混合余数浮点数值架构用于高吞吐量FPGA计算

A Hybrid Residue Floating Numerical Architecture with Formal Error Bounds for High Throughput FPGA Computation

Evolution Optiks R&D Inc.(进化光学研发有限公司)

Mostafa Darvishi

AI总结:

本文提出了一种结合无进位余数运算和指数基缩放的混合余数浮点数值架构,实现高吞吐量FPGA计算,具有宽动态范围和可预测误差,相比IEEE 754 FP32在性能和效率上有显著提升。

摘要:

浮点运算在FPGA平台上成本较高,由于宽数据路径、归一化和进位传播,这促使了替代数值表示法的发展,以提高吞吐量和效率。本文提出了混合余数浮点数值架构(HRFNA),这是一种完全指定的数值系统,结合了无进位余数运算和轻量级指数基缩放,以实现宽动态范围、可预测的误差行为和高效的FPGA实现。HRFNA具有严谨的数学基础:混合数字空间被正式定义,算术和归一化的正确性被证明,并且导出了显式的绝对和相对误差界限,将舍入限制在不频繁的归一化事件中。提出了一种完整的FPGA微架构,具有深度流水线的模运算、指数管理以及基于CRT的归一化引擎,在稳态操作下维持一个周期的启动间隔。在点积、密集矩阵乘法和迭代龙格-库塔微分方程求解器的应用级评估中,展示了在长计算序列中稳定的数值行为。在Xilinx Zynq UltraScale Plus ZCU104上实现后,HRFNA相比IEEE 754 FP32基准,实现了最高2.4倍的吞吐量,38至55%的LUT减少,以及最高1.9倍的能量效率提升,同时保持了有界的数值误差。比较分析显示,HRFNA占据了一个此前未被探索的设计点,介于数值稳定性、动态范围和硬件效率之间,使其非常适合FPGA为中心的科学和CAD相关计算。

英文摘要:

Floating point arithmetic is costly on FPGA platforms due to wide datapaths, normalization, and carry propagation, motivating alternative numerical representations that improve throughput and efficiency. This paper presents the Hybrid Residue Floating Numerical Architecture (HRFNA), a fully specified numerical system that combines carry-free residue arithmetic with lightweight exponent-based scaling to achieve wide dynamic range, predictable error behavior, and efficient FPGA implementation. HRFNA is developed with a rigorous mathematical foundation: the hybrid number space is formally defined, correctness of arithmetic and normalization is proven, and explicit absolute and relative error bounds are derived, confining rounding to infrequent normalization events. A complete FPGA microarchitecture is presented, featuring deeply pipelined modular arithmetic, exponent management, and a CRT-based normalization engine that sustains an initiation interval of one cycle under steady-state operation. Application-level evaluation on dot products, dense matrix multiplication, and iterative Runge-Kutta ODE solvers demonstrates stable numerical behavior over long computation sequences. Implemented on a Xilinx Zynq UltraScale Plus ZCU104, HRFNA achieves up to 2.4 times higher throughput, 38 to 55 percent LUT reduction, and up to 1.9 times energy efficiency improvement compared to IEEE 754 FP32 baselines, while maintaining bounded numerical error. Comparative analysis shows that HRFNA occupies a previously unexplored design point between numerical stability, dynamic range, and hardware efficiency, making it well suited for FPGA-centric scientific and CAD-relevant computation.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑