Accuracy–Performance Trade-offs in Ozaki-Inspired FP16 Matrix Multiplication on Qualcomm Mobile Hardware

Go to the profile of Asli Elif Ozgur
Asli Elif Ozgur
Student, Durham University
Like
Go to the profile of Hannah Bush

Share this post

Choose a social network to share with, or copy the URL to share elsewhere

Share with...

This is a representation of how your post may appear on social media. The actual post will vary between social networks

Preview

Open

I investigated the accuracy and performance trade-offs of FP16 matrix multiplication on Qualcomm mobile hardware. I found that a two-component, Ozaki-inspired method reduced numerical error by around 20%, although its runtime cost increased with matrix size. The study highlights both the potential and limitations of using low-precision arithmetic for efficient mobile computing. I am pleased to share the full research paper.

Go to the profile of Asli Elif Ozgur
Asli Elif Ozgur
Student, Durham University

Hi, I'm Asli, a first-year Computer Science student at Durham University. I'm excited to be spending this summer working on my research project and further developing my skills. I look forward to progressing through the programme, meeting new people, and learning from the experience along the way.

Cookie Settings

We and selected partners, use cookies or similar technologies as specified in the cookie policy and privacy policy.

You can consent to the use of such technologies by closing this notice.

Cookie Control

Customise your preferences for any tracking technology

The following allows you to customize your consent preferences for any tracking technology used to help us achieve the features and activities described below. To learn more about how these trackers help us and how they work, refer to the cookie policy. You may review and change your preferences at any time.

Cookie policy Privacy policy
Always active, cannot be turned off
More information

These trackers are used for activities that are strictly necessary to operate or deliver the service you requested from us and, therefore, do not require you to consent.

More information

These trackers help us to deliver personalized marketing content and to operate, serve and track ads.

More information

These trackers help us to deliver personalized marketing content to you based on your behaviour and to operate, serve and track social advertising.

More information

These trackers help us to measure traffic and analyze your behaviour with the goal of improving our service.

More information

These trackers help us to provide a personalized user experience by improving the quality of your preference management options, and by enabling the interaction with external networks and platforms.