SIMD AVX2 vs SSE4.2 GEMM Matrix Multiplication Performance: Vector Operations and MLP Inference on Intel Alder Lake

Tags: benchmarking, simd, research, opensource

Summary

This dataset characterizes the performance of SIMD-accelerated compute kernels in the Kazkade zero-copy columnar analytics engine. We compare AVX2, SSE4.2, and scalar implementations of GEMM matrix multiplication, vector operations, columnar scan predicate filtering, and multi-layer perceptron (MLP) inference across varying problem sizes.

Methodology

Hardware: Intel Core i7-1260P (AVX2, 8 cores, 12 MB L3 cache). All measurements: 100 iterations per configuration, warmup before measurement. Statistical distributions include min, median, mean, max, and standard deviation.

Key Results

• AVX2 GEMM at 1024x1024 achieves 33.8 GFLOPS ??? 9x faster than scalar (3.7 GFLOPS)
• SSE4.2 achieves 16.8 GFLOPS at the same size ??? 2x slower than AVX2
• Vector dot product: 10.9 GB/s with AVX2 vs 2.5 GB/s scalar
• MLP inference (3-layer, 128 neurons): 55,652 ops/sec at batch size 64
• Columnar scan on 10M rows: 78M rows/sec vs 21M rows/sec row-wise

Data Files

• stat_gemm.csv: GEMM performance across 3 implementations and 5 matrix sizes
• stat_vector_ops.csv: Vector operation throughput
• stat_mlp_inference.csv: MLP inference latency
• kazkade_cli_bench.txt: Raw CLI benchmark output

Dataset: https://doi.org/10.7910/DVN/YMJKOG