University of California

Knuedge, Inc.

The Intellisis Corporation

Friday Harbor LLC

Michael Florea

Douglas A Palmer

Jerome Vincent Coffin

Silvestre Yrra

A router that requests a reservation for an egress port prior to dequeuing a received packet. A reservation is granted only if there is space on the egress port for at least a maximum size packet. An ingress processor requests allocation of a packet buffer. An allocator grants the packet buffer, but if there are fewer than a threshold number of buffers available, the ingress processor will not accept the grant unless the received packet is to be routed to a port inside the device comprising the router. This conserves the packet buffer(s) for packets destined for locations within the device. After a reservation is obtained and a packet buffer has been accepted, the ingress processor begins dequeuing a received packet from an ingress port queue to the buffer, and provides an identifier of the buffer to an egress processor. The identifier is enqueued by the egress processor. After the identifier is dequeued, the egress processor copies the packet from the buffer to an egress queue and releases the buffer. The buffer can be released prior to completion of the data being copied.

Packet router buffer management

A neural processing engine may perform processing within a neural processing system and/or artificial neural network. The neural processing engine may be configured to effectively and efficiently perform the type of processing required in implementing a neural processing system and/or an artificial neural network. This configuration may facilitate such processing with neural processing engines having an enhanced computational density and/or processor density with respect to conventional processing units.

Neural processing engine and architecture using the same

A router requests a reservation for an egress port prior to dequeuing data from an ingress port data queue. A request is also made for the allocation of a buffer. After the reservation is received and a buffer is allocated, the data is copied the ingress port data queue to the buffer, and an identifier of the buffer is enqueued to an identifier queue of the egress port. After the identifier is dequeued, the data is copied from the buffer to an egress data queue of the egress port, and the buffer is released for reallocation. The buffer can be released prior to completion of the data being copied from the buffer. The queues associated with the egress port determine whether their depths equal or exceed a threshold associated with the respective egress queues. If one or more of the depths does equal or exceed the associated threshold, the granting of reservations for the egress port are postponed. The thresholds are each set to less than full capacity, so as to indicate that the respective queue is congested, rather than to indicate that they have no space available.

Flow control through packet router

Systems, devices, and techniques for routing packets are described. A described router includes ingress ports to receive packets; egress ports; ingress switch fabric coupled with the ingress ports; egress switch fabric coupled with the egress ports; floating buffers coupled between the ingress switch fabric and the egress switch fabric; and a controller. The controller can be configured to receive a packet via an ingress port, determine an egress port based on the packet's destination address, acquire a floating buffer, send to the egress port a buffer identifier corresponding to the acquired floating buffer, operate the ingress switch fabric to establish a first pathway between the acquired floating buffer and the ingress port to write the packet to the buffer, and operate the egress switch fabric to establish a second pathway between the acquired floating buffer and the egress port to write from the buffer to the egress port.

Router path selection and creation in a single clock cycle

The subject matter disclosed herein provides methods, apparatus, and articles of manufacture for neural-based processing. In one aspect, there is provided a method. The method may include reading, from a first memory, context information stored based on at least one connection value; reading, from a second memory, an activation value matching the at least one connection value; sending, by a first processor, the context information and the activation value to at least one of a plurality of microengines to configure the at least one microengine as a neuron; and generating, at the at least one microengine, a value representative of an output of the neuron. Related apparatus, systems, methods, and articles are also described.

Growing community of inventors

Lake Forest, CA, United States of America

Michael Florea