Tensor Core Matrix Multiplication on the NVIDIA T4: A WMMA FP16 GEMM with Double Buffering, Compared Against cuBLAS

Matrix multiplication (GEMM) is the main cost in deep-learning workloads, and Tensor Cores are the hardware NVIDIA GPUs provide to speed it up. This report describes a from-scratch CUDA GEMM that uses the WMMA API with FP16 inputs and FP32 accumulation, in two forms: a baseline with one shared-memory buffer and a double-buffered version that overlaps loading the next tile with computing the current one. Both were verified against cuBLAS and benchmarked on a Tesla T4 (Turing, sm_75) in Google Colab at three matrix sizes. The kernels reach about 63-67% of cuBLAS throughput. On this GPU, double buffering gave no measurable gain over the single-buffer baseline. The report explains the design, states that the T4 has no hardware asynchronous copy (cp.async), gives a likely reason for the flat result, and lists what would be tried next. Source code: https://github.com/19YearsOlds/61.-Tensor-Core-GEMM-WMMA-FP16-on-NVIDIA-T4

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24331218
Primary Topic
Parallel Computing and Optimization Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Tensor Core Matrix Multiplication on the NVIDIA T4: A WMMA FP16 GEMM with Double Buffering, Compared Against cuBLAS

Syed Tarteel Ejaz
Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
article

Tensor Core Matrix Multiplication on the NVIDIA T4: A WMMA FP16 GEMM with Double Buffering, Compared Against cuBLAS

Syed Tarteel Ejaz
article en

Abstract

Matrix multiplication (GEMM) is the main cost in deep-learning workloads, and Tensor Cores are the hardware NVIDIA GPUs provide to speed it up. This report describes a from-scratch CUDA GEMM that uses the WMMA API with FP16 inputs and FP32 accumulation, in two forms: a baseline with one shared-memory buffer and a double-buffered version that overlaps loading the next tile with computing the current one. Both were verified against cuBLAS and benchmarked on a Tesla T4 (Turing, sm_75) in Google Colab at three matrix sizes. The kernels reach about 63-67% of cuBLAS throughput. On this GPU, double buffering gave no measurable gain over the single-buffer baseline. The report explains the design, states that the T4 has no hardware asynchronous copy (cp.async), gives a likely reason for the flat result, and lists what would be tried next. Source code: https://github.com/19YearsOlds/61.-Tensor-Core-GEMM-WMMA-FP16-on-NVIDIA-T4

Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon