System and methods for mesh architecture for high bandwidth multicast and broadcast network

A hardware-based programmable mesh architecture to support high bandwidth multicast and broadcast for machine learning (ML), comprising:

• a plurality of processing blocks arranged in a two-dimensional array of a plurality of rows and columns;
• a plurality of routers associated with the plurality of processing blocks;
• and
• a plurality of bi-directional linkages that interconnect the plurality of routers of the processing blocks in the mesh architecture;
• wherein the mesh architecture is configured to
• i. accept a first data packet having an associated destination mask transferred from an external source at one processing block of the plurality of processing blocks, wherein the first data packet is to be routed to a plurality of destination processing blocks in the mesh structure as set in the destination mask;
• ii. route the first data packet vertically in Y direction to the same row as a first destination processing block according to an order set in the destination mask under a Y-X routing scheme;
• iii. route the first data packet horizontally in X direction to the first destination processing block when a second/next destination processing block is at the same row of the first/current destination processing block;
• when the second/next destination processing block is at a different row from the first/current destination processing block,
• iv. replicate the first data packet to create a replicated second data packet;
• v. route the first data packet horizontally in X direction to the first destination processing block;
• vi. route the replicated second data packet vertically in Y direction to the same row as the second destination processing block;
• vii. repeat steps iii to vi until the data packet and/or its replicated data packets have reached all destination processing blocks set in the destination mask.

A hardware-based programmable mesh architecture to support high bandwidth multicast and broadcast for machine learning (ML), comprising:

• a plurality of processing blocks arranged in a multi-dimensional array of a plurality of dimensions;
• a plurality of routers associated with the plurality of processing blocks;
• and
• a plurality of bi-directional linkages that interconnect the plurality of routers of the processing blocks in the mesh architecture;
• wherein the mesh architecture is configured to
• i. accept a first data packet having an associated destination mask transferred from an external source at one processing block of the plurality of processing blocks, wherein the first data packet is to be routed to a plurality of destination processing blocks in the mesh structure as set in the destination mask;
• ii. route the first data packet in a first direction across a first set of processing blocks according to an order set in the destination mask under a routing scheme;
• iii. route the first data packet in a second direction across a second set of processing blocks to a first destination processing block when a second/next destination processing block is in the same dimension of the first/current destination processing block;
• when the second/next destination processing block is at a different dimension from the first/current destination processing block,
• iv. replicate the first data packet to create a replicated second data packet;
• v. route the first data packet in the second direction to the first destination processing block;
• vi. route the replicated second data packet in a third direction across a third set of processing blocks;
• vii. repeat steps ill to vi until the data packet and/or its replicated data packets have reached all destination processing blocks set in the destination mask.

A method to support high bandwidth multicast and broadcast for machine learning (ML), comprising:

• i. accepting a first data packet having an associated destination mask transferred from an external source at one processing block of a plurality of processing blocks arranged in a two-dimensional array of a plurality of rows and columns of a programmable mesh architecture, wherein the first data packet is to be routed to a plurality of destination processing blocks in the mesh structure as set in the destination mask;
• ii. routing the first data packet horizontally in X direction to the same column as a first destination processing block according to an order set in the destination mask under a X-Y routing scheme;
• iii. routing the first data packet vertically in Y direction to the first destination processing block when a second/next destination processing block is at the same column of the first/current destination processing block;
• when the second/next destination processing block is at a different column from the first/current destination processing block,
• iv. replicating the first data packet to create a replicated second data packet;
• v. routing the first data packet vertically in Y direction to the first destination processing block;
• vi. routing the replicated second data packet horizontally in X direction to the same column as the second destination processing block;
• vii. repeating steps lii to vi until the data packet and/or its replicated data packets have reached all destination processing blocks set in the destination mask.

A method to support high bandwidth multicast and broadcast for machine learning (ML), comprising:

• i. accepting a first data packet having an associated destination mask transferred from an external source at one processing block of a plurality of processing blocks arranged in a two-dimensional array of a plurality of rows and columns of a programmable mesh architecture, wherein the first data packet is to be routed to a plurality of destination processing blocks in the mesh structure as set in the destination mask;
• ii. routing the first data packet vertically in Y direction to the same row as a first destination processing block according to an order set in the destination mask under a Y-X routing scheme;
• iii. routing the first data packet horizontally in X direction to the first destination processing block when a second/next destination processing block is at the same row of the first/current destination processing block;
• when the second/next destination processing block is at a different row from the first/current destination processing block,
• iv. replicating the first data packet to create a replicated second data packet;
• v. routing the first data packet horizontally in X direction to the first destination processing block;
• vi. routing the replicated second data packet vertically in Y direction to the same row as the second destination processing block;
• vii. repeating steps ill to vi until the data packet and/or its replicated data packets have reached all destination processing blocks set in the destination mask.

A method to support high bandwidth multicast and broadcast for machine learning (ML), comprising:

• i. accepting a first data packet having an associated destination mask transferred from an external source at one processing block of a plurality of processing blocks arranged in a multi-dimensional array of a plurality of dimensions of a programmable mesh architecture, wherein the first data packet is to be routed to a plurality of destination processing blocks in the mesh structure asset in the destination mask;
• ii. routing the first data packet in a first direction across a first set of processing blocks according to an order set in the destination mask under a routing scheme;
• iii. routing the first data packet in a second direction across a second set of processing blocks to a first destination processing block when a second/next destination processing block is in the same dimension of the first/current destination processing block;
• when the second/next destination processing block is at a different dimension from the first/current destination processing block,
• iv. replicating the first data packet to create a replicated second data packet;
• v. routing the first data packet in the second direction to the first destination processing block;
• vi. routing the replicated second data packet in a third direction across a third set of processing blocks;
• vii. repeating steps iii to vi until the data packet and/or its replicated data packets have reached all destination processing blocks set in the destination mask.