Fast Matrix Multiplication for AI

Projektinformationen

FMMF-AI

ID Finanzhilfevereinbarung: 101113120

DOI

10.3030/101113120

Projekt abgeschlossen

EK-Unterschriftsdatum 23 März 2023

Startdatum 1 April 2023

Enddatum 30 September 2024

Finanziert unter

• European Research Council (ERC)

Gesamtkosten

Keine Daten

EU-Beitrag

€ 150 000,00

Koordiniert durch

THE HEBREW UNIVERSITY OF JERUSALEM
Israel

Periodic Reporting for period 1 - FMMF-AI (Fast Matrix Multiplication for AI)

Berichtszeitraum: 2023-04-01 bis 2024-09-30

Zusammenfassung vom Kontext und den Gesamtzielen des Projekts

• Matrix multiplication computation consumes huge amount of resources: computing time and energy, primarily in AI applications.
• The industry has recognized the need for faster and more energy-efficient matrix multiplication with state-of-the-art solutions in software (e.g. DGEMM of Intel's math kernel library (MKL) for CPU and NVIDIA's CUDA for GPU) and hardware (e.g. Google's TPU and Intel / Habana labs accelerator). Unfortunately, all present solutions employ a wasteful cubic-time algorithm.
• We have developed speedup for matrix multiplication, that can accelerate computations by factor x2-x10, offering costs-saving, time-saving, and energy-saving. Our solution can be implemented in software and hardware. In a preliminary benchmarking study, we outperformed Intel by a factor of about x2: our code vs. Intel's DGEMM tested on their Xeon E5-2680 CPU.
• The novel developments of Prof. Oded Schwartz and his strong team are based on many years of research, and are protected by several patents. The funds are requested to pursue business opportunity.

Arbeit, die ab Beginn des Projekts bis zum Ende des durch den Bericht erfassten Berichtszeitraums geleistet wurde, und die wichtigsten bis dahin erzielten Ergebnisse

We have implemented dense matrix-matrix multiplication based on the alternative basis method. Tests were performed on various CPU and GPU hardware settings. Inputs included random matrices of various dimensions, square and rectangular. Our code proved to outperform vendor-tuned libraries, such as MKL for Intel's processors, and Cuda/Cublas for Nvidia's GPUs.
Project includes incorporation of the new implementations into GenAI workflows.
Interaction with potential customers and potential private investors has been initiated.

Fortschritte, die über den aktuellen Stand der Technik hinausgehen und voraussichtliche potenzielle Auswirkungen (einschließlich der bis dato erzielten sozioökonomischen Auswirkungen und weiter gefassten gesellschaftlichen Auswirkungen des Projekts)

Our code proved to outperform vendor-tuned libraries, such as MKL for Intel's processors, and Cuda/Cublas for Nvidia's GPUs.
A new patent has been submitted.