Residue number system in a photonic matrix accelerator

Patent metadata

Field	Value
Publication number	US-12307219-B2
Application number	US-202318496494-A
Country	US
Kind code	B2
Filing date	Oct 27, 2023
Priority date	Dec 7, 2020
Publication date	May 20, 2025
Grant date	May 20, 2025

Abstract

A photonic processor uses light signals and a residue number system (RNS) to perform calculations. The processor sums two or more values by shifting the phase of a light signal with phase shifters and reading out the summed phase with a coherent detector. Because phase winds back every 2π radians, the photonic processor performs addition modulo 2π. A photonic processor may use the summation of phases to perform dot products and correct erroneous residues. A photonic processor may use the RNS in combination with a positional number system (PNS) to extend the numerical range of the photonic processor, which may be used to accelerate homomorphic encryption (HE)-based deep learning.

First claim

What is claimed is:

1. A photonic processor for representing numbers using a redundant residue number system (RRNS) that includes a set of moduli, the set of moduli including n moduli and k redundant moduli, the photonic processor comprising:
at least one optical transmitter configured to, for each modulus of the set of moduli:
provide a first value associated with the modulus; and
provide a second value associated with the modulus;

at least one phase shifter configured to, for each modulus of the set of moduli:
shift, by each of the first value and the second value, a phase of a light signal; and

at least one receiver configured to, for each modulus of the set of moduli:
detect the shifted phase of the light signal; and
output the detected phase of the light signal as representing a residue of a sum of the first value and the second value, the residue being associated with the modulus,

wherein the photonic processor is further configured to use a residue associated with a modulus of the k redundant moduli to detect and/or correct an error of a residue associated with a modulus of the n moduli,
wherein the photonic processor is configured to detect and/or correct the error by:
forming a plurality of sets of moduli, each set of moduli comprising p moduli;
determining a plurality of representations by, for each set of moduli of the plurality of sets of moduli, converting a set of residues associated with the set of moduli to a representation of the set of residues;
determining a correct representation using the plurality of representations.

2. The photonic processor of claim 1 , wherein determining the correct representation using the plurality of representations comprises:
selecting a representation of the plurality of representations that occurs most often in the plurality of representations.

3. The photonic processor of claim 1 , wherein the representation of the set of residues comprises an integer.

4. The photonic processor of claim 1 , wherein converting the set of residues associated with the set of moduli to the representation of the set of residues comprises:
converting the set of residues associated with the set of moduli to the representation without using Chinese Remainder Theorem (CRT).

5. The photonic processor of claim 1 , wherein p is equal to n.

6. The photonic processor of claim 1 , wherein the photonic processor is configured to detect and/or correct the error by:
providing a first residue associated with a redundant modulus of the k redundant moduli;
using a residue associated with a modulus of the n moduli, computing a second residue associated with the redundant modulus of the k redundant moduli;
comparing the first residue with the second residue; and
determining that there is an erroneous residue when the first residue does not match the second residue.

7. The photonic processor of claim 6 , wherein the photonic processor is further configured to detect and/or correct the error by:
determining whether the erroneous residue is associated with a modulus of the n moduli, or a redundant modulus of the k redundant moduli.

8. The photonic processor of claim 7 , wherein the photonic processor is further configured to detect and/or correct the error by:
responsive to determining that the erroneous residue is associated with a modulus of the n moduli:
computing an index using the first residue and the second residue each associated with the redundant modulus of the k redundant moduli;
using the index, looking up a correction factor in an error correction table; and
computing a corrected residue using the erroneous residue and the correction factor.

9. The photonic processor of claim 1 , wherein the photonic processor is configured to:
responsive to determining that there is a number of errors that is greater than an error correction threshold, perform maximum likelihood (ML) decoding to determine a set of corrected residues that are closest to a set of residues associated with the detected phases of light.

10. A method of using a photonic processor to represent numbers using a redundant residue number system (RRNS) that includes a set of moduli, the set of moduli including n moduli and k redundant moduli, the method comprising:
for each modulus of the set of moduli:
providing a first value associated with the modulus;
providing a second value associated with the modulus;
shifting, by each of the first value and the second value, a phase of a light signal;
detecting the shifted phase of the light signal; and
outputting the detected phase of the light signal as representing a residue of a sum of the first value and the second value, the residue being associated with the modulus; and

using a residue associated with a modulus of the k redundant moduli to detect and/or correct an error of a residue associated with a modulus of the n moduli,
wherein detecting and/or correcting the error comprises:
forming a plurality of sets of moduli, each set of moduli comprising p moduli;
determining a plurality of representations by, for each set of moduli of the plurality of sets of moduli, converting a set of residues associated with the set of moduli to a representation of the set of residues;
determining a correct representation using the plurality of representations.

11. The method of claim 10 , wherein determining the correct representation using the plurality of representations comprises:
selecting a representation of the plurality of representations that occurs most often in the plurality of representations.

12. The method of claim 10 , wherein the representation of the set of residues comprises an integer.

13. The method of claim 10 , wherein converting the set of residues associated with the set of moduli to the representation of the set of residues comprises:
converting the set of residues associated with the set of moduli to the representation without using Chinese Remainder Theorem (CRT).

14. The method of claim 10 , wherein p is equal to n.

15. The method of claim 10 , wherein detecting and/or correcting the error comprises:
providing a first residue associated with a redundant modulus of the k redundant moduli;
using a residue associated with a modulus of the n moduli, computing a second residue associated with the redundant modulus of the k redundant moduli;
comparing the first residue with the second residue; and
determining that there is an erroneous residue when the first residue does not match the second residue.

16. The method of claim 15 , wherein detecting and/or correcting the error further comprises:
determining whether the erroneous residue is associated with a modulus of the n moduli, or a redundant modulus of the k redundant moduli.

17. The method of claim 16 , wherein detecting and/or correcting the error further comprises:
responsive to determining that the erroneous residue is associated with a modulus of the n moduli:
computing an index using the first residue and the second residue each associated with the redundant modulus of the k redundant moduli;
using the index, looking up a correction factor in an error correction table; and
computing a corrected residue using the erroneous residue and the correction factor.

18. The method of claim 10 , wherein detecting and/or correcting the error further comprises:
responsive to determining that there is a number of errors that is greater than an error correction threshold, perform maximum likelihood (ML) decoding to determine a set of corrected residues that are closest to a set of residues associated w

Assignees

• Lightmatter Inc

Inventors

• Hein Eric
• Basumalik Ayon
• Harris Nicholas C
• Bunandar Darius
• Demirkiran Cansu

Classifications

• G06E1/00

  Devices for processing exclusively digital data · CPC title

• H04L9/3093

  involving Lattices or polynomial equations, e.g. NTRU scheme · CPC title

• H04L9/008

  involving homomorphic encryption · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/0675

  using electro-optical, acousto-optical or opto-electronic means · CPC title