Skip to content
Conference Open access

High-Performance GEMV on AMD Zen 3 via Architecture-Specific AVX2/FMA Optimization

Aug 2026 · Conference on Computer Science and Information Systems · Vol 47, pp. 357-364 · 0 citations · 29 references
Computer Science

TL;DR

A hand-optimized GEMV kernel for AMD Zen 3, implemented using AVX2 (Advanced Vector Extensions 2) and FMA (fused multiply-add) intrinsics is presented, achieving up to 1.37× speedup over OpenBLAS and 1.20× over AOCL-BLAS.

Abstract

The General Matrix-Vector Multiply (GEMV) is a fundamental Level-2 BLAS operation appearing throughout scientific computing and numerical linear algebra. GEMV is inherently memory-bound, making effective cache use the dominant optimization factor.We present a hand-optimized GEMV kernel for AMD Zen 3, implemented using AVX2 (Advanced Vector Extensions 2) and FMA (fused multiply-add) intrinsics. Our single-threaded kernel (ZenGEMV) processes 4 rows simultaneously with dual accumulators per row, creating 8 independent FMA chains that fully saturate the Zen 3 pipeline. Benchmarked against OpenBLAS 0.3.31 and AOCL-BLAS 5.2.0 on an AMD Ryzen 5 5600, ZenGEMV achieves up to 1.37× speedup over OpenBLAS and 1.20× over AOCL-BLAS. Our parallel variant (ZenGEMV-P) achieves up to $3.22 \times$ speedup over OpenBLAS on small matrices.We analyse why dispatch overhead, suboptimal accumulator strategies, and generic design choices in general-purpose BLAS libraries leave performance gains available in architecture-specific workloads.

Read PDF

Similar papers

#artificial intelligence Preprint Oct 2026

Waypoint: Analytically Modeling GEMM for CPUs

High-performance General Matrix Multiplication (GEMM) implementations rely on breaking down GEMM into carefully selected tiles to achieve peak performance. We present Waypoint, a new analytical model to automatically identify the optimal tile size for a given hardware and input matrix shape. Waypoint estimates the achi...

N. Srinidhi, Sunanda Somwase, Hemant Kumar Chandrakant Chandavar et al. · 0 citations
Book Open access Sep 2026

TileGEMM: Boosting the Performance of GEMM on AMX-Powered CPUs by Exploiting Data Reuse

TileGEMM is proposed, a high-performance GEMM implementation on AMX that systematically enhances data reuse across the memory hierarchy, and achieves average speedups of 3.27 × and 1.96 × over AVX-512-based implementations TVM and MKL, respectively.

Kang-Kang Chen, Hua-You Su, Meng-Han Jia et al. · 0 citations
Preprint Sep 2026

Parallelizing the Factorial Space: Multi-Core OpenMP Scaling and Scalable SIMD Acceleration of the Steinhaus-Johnson-Trotter Algorithm via Dual-Lane AVX2 Execution

A high-performance SIMD acceleration framework for the Steinhaus-Johnson-Trotter algorithm, targeted at modern x86-64 architectures using the AVX2 instruction set, extended into a highly concurrent environment via OpenMP using a localized mathematical state decoder and macro-period loop scheduling.

S. Mel'nikov · 0 citations
#machine learning Preprint Oct 2026

Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs

Meta deploys a heterogeneous fleet comprising NVIDIA GPUs, AMD GPUs, and proprietary MTIA accelerators to reduce costs and mitigate vendor lock-in. However, the extensive history of NVIDIA-specific optimizations hinders performance portability to other hardware. This paper presents our methods for optimizing alternativ...

Hao-Yu Zhang, Arshad Rahman, Adnan Aziz et al. · 0 citations
Sep 2026

Rethinking Sparse Formats for RISC-V: A Hierarchical Approach to High-Performance SpMV

A novel modification of the CSR format, Hierarchical CSR (HCSR), is proposed, which enhances SpMV performance on RISC-V processors and achieves the shortest execution time among all considered formats across a broad class of sparse matrices.

A. Pirova, Anastasia Vodeneeva, K. Kovalev et al. · 0 citations
#machine learning Preprint Sep 2026

Programming AMD XDNA NPUs with Open-source Compiler Tools: A FlashAttention Case Study

This work compares four reference designs on XDNA 1 and XDNA 2: one runs each operator separately, two stream between operators on chip, and one fuses all three attention stages into a single kernel, which covers twelve LLM configurations, from BERT to DeepSeek, up to 128K tokens.

Erwei Wang, Ephrem C. Wu, V. J. Jung et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.