Work

qgemv-roofline: a quantized GEMV kernel ladder measured against its own bandwidth roof

Year: 2026

Type: software

Abstract: Twelve CUDA kernels and a Triton implementation of the batch-1 LLM decode GEMV over ggml q8_0 and q4_0 weights, measured on five NVIDIA GPUs spanning sm_60 to sm_89 against a bandwidth ceiling the harness measures for itself.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.22164007

Author: Jacob, Manu Nicholas

Language: English

Topic: Parallel Computing and Optimization Techniques

Subfield: Hardware and Architecture

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7204643456

Cites:

Cited by:

Related to: 10