US 12,217,147 B2Granted

Advanced wavelet filtering for accelerated deep learning

AssigneeCerebras Systems Inc.

InventorsMichael Morrison, Michael Edwin James, Sean Lie, Srikanth Arekapudi, Gary R. Lauterbach

Life of this patent

Estimating the term…

Loading patent…

Abstract

Techniques in wavelet filtering for advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element comprises a compute element to execute programmed instructions using the data and a router to route the wavelets in accordance with virtual channel specifiers. Each processing element is enabled to perform local filtering of wavelets received at the processing element, selectively, conditionally, and/or optionally discarding zero or more of the received wavelets, thereby preventing further processing of the discarded wavelets. The wavelet filtering is performed by one or more configurable wavelet filters operable in various modes, such as counter, sparse, and range modes.

Claims32

Dependent claims

A method comprising:

communicating packets between a plurality of processing elements coupled via a fabric, wherein each of the communicated packets comprises a respective color field and a respective type field;
receiving, from the fabric, a particular one of the communicated packets in a particular one of the processing elements, wherein the color field of the particular packet is a particular color value and the type field of the particular packet is a particular type value indicating a mutually exclusive one of a plurality of types;
managing a wavelet filter of the particular processing element, wherein the wavelet filter is enabled to retain state comprising a filtered color;
and
queueing information of the particular packet to enable execution, by the particular processing element, of an instruction that uses the queued information, wherein the queueing is conditional based on one or more queueing criteria, a first one of the queueing criteria is that the particular color value and the filtered color are different, and a second one of the queueing criteria is that the particular type value indicates a particular one of the types.

2 more elements

A method comprising:

communicating packets between a plurality of processing elements coupled via a fabric, wherein each of the communicated packets comprises a respective color field and a respective type field;
receiving, from the fabric, a particular one of the communicated packets in a particular one of the processing elements, wherein the color field of the particular packet is a particular color value and the type field of the particular packet is a particular type value indicating a mutually exclusive one of a plurality of types;
managing a wavelet filter of the particular processing element, wherein the wavelet filter is enabled to retain state comprising a filtered color;
queueing information of the particular packet to enable execution, by the particular processing element, of an instruction that uses the queued information, wherein the queueing is conditional based on one or more queueing criteria and a first one of the queueing criteria is that the particular color value and the filtered color are different, and a second one of the queueing criteria is that the particular type value indicates a particular one of the types;
and
updating a counter, wherein:
the updating is conditional based on one or more updating criteria and a first one of the updating criteria is that the particular color value and the filtered color are identical, and a second one of the updating criteria is that the particular type value indicates the particular type,
the updating is responsive to the receiving, and
the state further comprises the counter.

6 more elements

Dependencies are inferred from the claim text.

Facts

Application: 17/764,453Filed 15 Oct 2020 · us provisional application
Priority: 16 Oct 2019
Published: US 2022/0343136 A127 Oct 2022
Granted: 4 Feb 2025
Term ends: Not estimatedThe source records don’t support an estimate.

Classification15 codes

G06N 3/04
Neural networks: architecture, e.g. interconnection topology
G06F 9/3887
Using a plurality of independent parallel functional units: controlled by a single instruction for multiple data lanes [SIMD]
G06F 9/3888
Using a plurality of independent parallel functional units: controlled by a single instruction for multiple threads [SIMT] in parallel
G06F 13/42
Information transfer, e.g. on bus: bus transfer protocol, e.g. handshake; Synchronisation