Proposal for a High Precision Tensor Processing Unit

TL;DR

This paper proposes a new high-precision Tensor Processing Unit that leverages fractional arithmetic based on the residue number system to enhance performance on wide precision data operations.

Contribution

It introduces a novel TPU design utilizing fractional arithmetic with residue number systems for improved high-precision computation.

Findings

Achieves high performance comparable to Google's TPU

Supports wide precision data operations effectively

Demonstrates the feasibility of residue number system-based arithmetic

Abstract

This whitepaper proposes the design and adoption of a new generation of Tensor Processing Unit which has the performance of Google's TPU, yet performs operations on wide precision data. The new generation TPU is made possible by implementing arithmetic circuits which compute using a new general purpose, fractional arithmetic based on the residue number system.