# US 12,603,829 B2 · Network collective offload message chunking management

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,603,829 B2
- Document ID: US12603829B2
- Status: Granted
- Kind: B2
- Application number: 18/540,782
- Filed: 2023-12-14
- Published: 2024-06-20
- Granted: 2026-04-14
- Earliest claimed: 2022-12-14
- Assignees: Advanced Micro Devices, Inc.
- Applicants: Advanced Micro Devices, Inc.
- Inventors: Josiah I. Clark
- Web page: https://patentgrep.com/patent/US12603829B2

## CPC

- H04L 45/02: Topology update or discovery
- H04L 41/0823: characterised by the purposes of a change of settings, e.g. optimising configuration for enhancing reliability (for optimising operational conditions of wireless networks H04W24/02)
- H04L 41/12: Discovery or management of network topologies
- H04L 41/145: {involving simulating, designing, planning or modelling of a network}
- H04L 45/34: {Source routing}
- H04L 49/30: {Peripheral units, e.g. input or output ports}
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

The disclosed device can perform a collective operation on received datasets, and split the result into chunks in accordance with a chunking scheme. The device can also forward the chunks in accordance with a routing scheme that can direct chunks to appropriate nodes of a collective network. Various other methods, systems, and computer-readable media are also disclosed.

## Claims

### Claim 1

A method comprising:
receiving a plurality of datasets;
performing a collective operation of a collective network on the plurality of datasets to produce a result dataset;
split the result dataset into a plurality of chunks using a chunking scheme that correlates index values of the plurality of datasets to a plurality of destinations in a round-robin style; and
send the plurality of chunks to the plurality of destinations based on the chunking scheme and a topology corresponding to the collective operation that defines paths for chunks traversing the collective network such that corresponding chunks reach a same destination.

### Claim 2 (depends on 1)

The method of claim 1, wherein the topology defines connections between nodes of a collective network based on the collective operation such that the plurality of datasets are received from a prior level of the collective network and the plurality of chunks are sent to a subsequent level of the collective network.

### Claim 3 (depends on 1)

The method of claim 1, wherein a number of the plurality of datasets corresponds to a number of input ports and a number of the plurality of chunks corresponds to a number of a plurality of output ports.

### Claim 4 (depends on 1)

The method of claim 1, wherein the paths for chunks traversing the collective network ensures that chunks of matching vector indexes reach the same destination for the collective operation.

### Claim 5

A device comprising:
a plurality of output ports; and
a control circuit configured to:
perform a collective operation on a dataset to produce a result dataset;
split the result dataset into a plurality of chunks using a chunking scheme that correlates index values of the dataset to the plurality of output ports in a round-robin style; and
send the plurality of chunks through the plurality of output ports based on the chunking scheme.

### Claim 6 (depends on 5)

The device of claim 5, wherein a number of the plurality of chunks matches a number of the plurality of output ports such that each of the plurality of chunks are sent through a respective one of the plurality of output ports.

### Claim 7 (depends on 5)

The device of claim 5, wherein the device is part of collective network such that the plurality of output ports are connected to nodes of a subsequent level of the collective network.

### Claim 8 (depends on 7)

The device of claim 7, wherein the plurality of output ports are connected to the nodes of the subsequent level based on a topology that defines paths for chunks traversing the collective network to send each of the plurality of chunks to destinations defined by the topology such that corresponding chunks reach a same destination.

### Claim 9 (depends on 8)

The device of claim 8, wherein the topology is enforced using segment routing of the plurality of chunks.

### Claim 10 (depends on 8)

The device of claim 8, wherein the topology defines connections between nodes based on the collective operation.

### Claim 11 (depends on 5)

The device of claim 5, further comprising a plurality of input ports, wherein the control circuit is further configured to:
receive, through the plurality of input ports, a plurality of datasets; and
perform the collective operation on the plurality of datasets to produce the result dataset.

### Claim 12 (depends on 11)

The device of claim 11, wherein a number of the plurality of datasets matches a number of the plurality of input ports such that each of the plurality of datasets are received from a respective one of the plurality of input ports.

### Claim 13 (depends on 11)

The device of claim 11, wherein the device is part of collective network such that the plurality of input ports are connected to nodes of a prior level of the collective network.

### Claim 14 (depends on 13)

The device of claim 13, wherein the plurality of datasets corresponds to chunks sent from the nodes of the prior level of the collective network.

### Claim 15

A device comprising:
a plurality of input ports;
a plurality of output ports; and
a control circuit configured to:
receive, through the plurality of input ports, a plurality of datasets;
perform a collective operation on the plurality of datasets to produce a result dataset;
split the result dataset into a plurality of chunks using a chunking scheme that correlates index values of the plurality of datasets to the plurality of output ports in a round-robin style; and
send the plurality of chunks through the plurality of output ports based on the chunking scheme.

### Claim 16 (depends on 15)

The device of claim 15, wherein a number of the plurality of datasets matches a number of the plurality of input ports such that each of the plurality of datasets are received from a respective one of the plurality of input ports.

### Claim 17 (depends on 15)

The device of claim 15, wherein a number of the plurality of chunks matches a number of the plurality of output ports such that each of the plurality of chunks are sent through a respective one of the plurality of output ports.

### Claim 18 (depends on 15)

The device of claim 15, wherein the device is part of collective network such that the plurality of input ports are connected to nodes of a prior level of the collective network and the plurality of output ports are connected to nodes of a subsequent level of the collective network.

### Claim 19 (depends on 18)

The device of claim 18, wherein the plurality of datasets corresponds to chunks sent from the nodes of the prior level of the collective network.

### Claim 20 (depends on 18)

The device of claim 18, wherein the plurality of input ports are connected to the nodes of the prior level and the plurality of output ports are connected to the nodes of the subsequent level based on a topology that defines paths for chunks traversing the collective network and the plurality of chunks are sent through the plurality of output ports to destinations defined by the topology such that corresponding chunks reach a same destination.