Accelerating exponential functions on Arm: a practical introduction to FEXPA

Modern workloads, from scientific computing to machine learning, rely heavily on exponential functions. When you implement activation functions such as Softmax or optimize numerical kernels, computing e^x can become a bottleneck. On Arm-based systems the FEXPA instruction is a powerful and underused feature. In this post, we explore how FEXPA works, why it matters, and how you can use it to unlock significant performance gains. 

The standard library implementation of the exponential function (expf) is accurate but scalar, so it does not optimize for throughput. The first level of optimization is to use SVE (Scalable Vector Extension), apply range reduction, and approximate the function with low-degree polynomials. 
This approach can improve performance by 1.5 to 4 times, depending on the hardware and vector length. 

The most effective implementations combine range reduction and table lookup.  FEXPA stands for Floating-point EXPonential Accelerator. It uses bit manipulation on floating-point values, leverages hardware lookup tables and constructs results by combining exponent and fraction components. This hybrid approach uses lower polynomial degree and fewer instructions, hence it provides up to 6 times speedup over the scalar function. Moreover, SME support for FEXPA lets you embed the exponential directly into the matrix computation path, which translates into higher throughput, lower power and bandwidth, and cleaner fusion with GEMM workloads. 

If your workload depends on exponential functions, FEXPA can improve performance, while maintaining accuracy. As AI and HPC workloads grow, these optimizations become more important. 

The following Learning Path explains each step.

Optimize exponential functions with FEXPA