International Business Machines Corporation

Globalfoundries Inc.

Yasuteru Kohda

Nobuyuki Ohba

Kohji Takano

Daiju Nakano

Yasunao Katayama

Chitra K Subramanian

Seiji Munetoh

Geoffrey W Burr

Akihiro Horibe

Kuniaki Sueoka

Shubham Jain

Milos Stanisavljevic

Zhang Gang

Efficient data layout and alignment techniques for effectively executing AI workloads in wide-vector accelerator systems are provided. In one aspect, a method for processing AI workloads includes: logically dividing a data vector into a hierarchy of segments and sub-segments with each of the segments including more than one of the sub-segments, wherein each of the sub-segments includes words, and each of the words includes data-bits; and physically mapping the data-bits such that the words belonging to a same given one of the sub-segments are mapped contiguously across all of the segments. An AI accelerator system is also provided.

Efficient data layout and alignment for wide-vector accelerator systems

An efficient pipelined implementation of digital scaling, offset and aggregation operation supports element-by-element programmable scale and offset factors. The method includes time-multiplexed parallel pipelining of a plurality of digital data words, each of the plurality of digital data words encoding an N-bit signed integer, from one of a plurality of receive-registers through a datapath that can either (1) store the plurality of digital data words directly in a dedicated first memory, (2) store the plurality of digital data words directly in a dedicated second memory, or (3) direct the plurality of digital data words into a parallel set of fused-multiply-add units. The method further includes multiplying each digital data word by a corresponding data-word retrieved from the dedicated first memory to form product data words and adding the product data words to a corresponding data-word retrieved from the dedicated second memory to form an output sum-and-product data words.

Special-purpose digital-compute hardware for efficient element-wise aggregation, scaling and offset

A method for arbitrating data transfer requests from a plurality of nodes includes specifying one or more nodes among the plurality of nodes, the one or more nodes satisfying a predetermined condition, and selecting, if two or more nodes are specified among the plurality of nodes, one node from the two or more nodes using priority information, the priority information indicating correspondence between the plurality of nodes and a plurality of priorities each assigned to one of the plurality of nodes, the correspondence changing so that the plurality of priorities are assigned equally to each of the plurality of nodes and high and low relations appear equally between pairs of priorities each assigned to a pair of nodes of the plurality of nodes.

Arbitration of data transfer requests

A system is provided for packet broadcasting in a mesh-interconnected multi-computer network having a plurality of routers. The system includes a plurality of arbiters, interconnected to each of the plurality of routers, configured to live-lock free arbiter between the two or more packet broadcast requests using a shared priority matrix, implemented by a binary matrix, that selects one of the two or more packet broadcast requests and includes a column for each of the plurality of routers. The shared priority matrix is shared amongst the plurality of arbiters and stores priority information determined from summing the matrix column values and relating to a correspondence between a plurality of packet broadcast requests, including the two or more packet broadcast requests, with respect to priorities assigned to each of the plurality of packet broadcast requests. Each of columns of the shared priority matrix corresponds to a respective one of the routers.

Packet broadcasting mechanism for mesh interconnected multi-computers

A method is provided for packet broadcasting in a mesh-interconnected multi-computer network having a plurality of routers interconnected to a plurality of arbiters. The method includes live-lock free arbitering, by each of the plurality of arbiters, between two or more packet broadcast requests using a shared priority matrix, implemented by a binary matrix, that selects one of the two or more packet broadcast requests and includes a column for each of the plurality of routers, the shared priority matrix being shared amongst the plurality of arbiters and storing priority information determined from summing the matrix column values and relating to a correspondence between a plurality of packet broadcast requests, including the two or more packet broadcast requests, with respect to priorities assigned to each of the plurality of packet broadcast requests. Each of the columns of the shared priority matrix corresponds to a respective one of the routers.

A method, system, and data structure for sending data in a network which includes a plurality of nodes is provided. The method includes sending the data from a source node to an intermediate node. The method also includes sending the data from the intermediate node to plural destination nodes within a portion of the network. The portion is defined based on at least one physical distance from the intermediate node.

Multi-step remote packet broadcasting/multicasting mechanism for cognitive systems

A method for routing data by a node from a plurality of nodes connected by links to form a network includes feeding, to a global bus, local traffic information regarding the node, the global bus being provided separately from the links and spread within the network. The method also includes obtaining, from the global bus, global traffic information regarding other nodes from the plurality of nodes. The method further includes switching a routing path of the data based on at least the local traffic information regarding the other nodes.

Routing of data in network

A chip intermediate body includes a semiconductor region including plural chip areas. The chip areas respectively are cut out as semiconductor chips. A cut region is provided along edges of the chip areas, the cut region being cut to cut out the semiconductor chips. A contact region is provided opposite to the chip areas across the cut region, the contact region being configured to be contacted by a probe of a test unit to test the chip areas, and electric wiring is provided continuously with the cut region to connect the chip areas and the contact region.

Wafer scale testing and initialization of small die chips

A method is provided for detour routing in a mesh network that includes nodes and bi-directional links connecting neighboring nodes. The method identifies any primary links useable to decrease a path distance from a source to a destination node. The method randomly selects one primary link for use, responsive to identifying more than one primary link that is non-defective. The method identifies any secondary links useable to decrease the path distance and orthogonal to a primary link, responsive to the primary links being defective. The method randomly selects one secondary link for use, responsive to identifying more than one secondary link that is non-defective. The method identifies any tertiary links useable to decrease the path distance and reverse in direction to a primary link, responsive to the secondary links being defective. The method sends a packet from the source to the destination node using the identified tertiary links.

Detour routing for mesh network

A method is provided for packet broadcasting in a mesh-interconnected multi-computer network having a plurality of routers and a plurality of arbiters. The method includes interconnecting each of the plurality of routers to each of the plurality of arbiters. The method further includes receiving, by each of the plurality of arbiters, two or more packet broadcast requests. The method also includes live-lock free arbitering, by each of the plurality of arbiters, between the two or more packet broadcast requests using a shared priority matrix that selects one of the two or more packet broadcast requests. The shared priority matrix is shared amongst the plurality of arbiters and stores priority information relating to a correspondence between a plurality of packet broadcast requests, including the two or more packet broadcast requests, with respect to priorities assigned to each of the plurality of packet broadcast requests.

A network system for routing data in a network includes a plurality of nodes each node from the plurality of nodes being configured to route the data, a plurality of links each link from the plurality of links connects two of the plurality of nodes to form the network, and a global bus provided separately from the plurality of links and spread within the network, wherein each node from the plurality of nodes feeds, to the global bus, traffic information regarding each node, obtains, from the global bus, traffic information regarding other nodes from the plurality of nodes, and switches a routing path of the data based on at least the traffic information regarding the other nodes.

Growing community of inventors

Kanagawa-ken, Japan

Yasuteru Kohda