# US 2025/0165285 A1 · Iterative system for massively parallel processing and methods for use therewith

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

- Number: US 2025/0165285 A1
- Document ID: US20250165285A1
- Status: Published application
- Kind: A1
- Application number: 18/901,057
- Filed: 2024-09-30
- Published: 2025-05-22
- Earliest claimed: 2023-11-16
- Assignees: BEIT Sp. z o.o.
- Applicants: BEIT Sp. z o.o.
- Inventors: Adam Szady; Witold Jarnicki
- Web page: https://patentgrep.com/patent/US20250165285A1

## CPC

- G06F 9/4881: {Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues}
- G06F 9/5066: {Algorithms for mapping a plurality of inter-dependent sub-tasks onto a plurality of physical CPUs (mappping at compile time, see G06F8/451)}
- G06F 9/5038: {considering the execution order of a plurality of tasks, e.g. taking priority or time dependency constraints into consideration (scheduling strategies G06F9/4881 and subgroups)}
- G06F 9/5072: {Grid computing}
- G06F 9/546: {Message passing systems or structures, e.g. queues}
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A computing system includes: a computational engine having a plurality of processing elements arrayed in accordance with a grid topology, wherein each of the plurality of processing elements is configured to perform one of a plurality of processing functions, and wherein the plurality of processing functions include a principal function, a kernel function, a routing function, and a reducer function; and a host processor having at least one processing circuit and at least one memory that stores operational instructions that, when executed by the at least one processing circuit, cause the at least one processor to: receive a task input; and control the computational engine in accordance with an iterative reduce-broadcast-apply process to generate a task output that is based on the task input.

## Claims

### Claim 1

A computing system comprising:
a computational engine having a plurality of processing elements arrayed in accordance with a grid topology, wherein each of the plurality of processing elements is configured to perform one of a plurality of processing functions, and wherein the plurality of processing functions include a principal function, a kernel function, a routing function, and a reducer function; and
a host processor having at least one processing circuit and at least one memory that stores operational instructions that, when executed by the at least one processing circuit, cause the at least one processing circuit to:
receive a task input; and
control the computational engine in accordance with an iterative reduce-broadcast-apply process to generate a task output that is based on the task input.

### Claim 2 (depends on 1)

The system of claim 1, wherein the iterative reduce-broadcast-apply process includes a produce operation, a reduce operation, a broadcast operation and an apply operation.

### Claim 3 (depends on 2)

The system of claim 2, wherein the produce operation is implemented by a subset of the plurality of processing elements performing corresponding kernel functions that operate in parallel to generate corresponding operation shards.

### Claim 4 (depends on 3)

The system of claim 3, wherein the reduce operation is implemented by a subset of the plurality of processing elements performing corresponding reducer functions that operate to receive two operation shards as input and produce a single operation shard as output.

### Claim 5 (depends on 4)

The system of claim 4, wherein a subset of the plurality of processing elements performing corresponding router functions operate to perform message routing between others of the plurality of processing elements.

### Claim 6 (depends on 3)

The system of claim 3, wherein the iterative reduce-broadcast-apply process further includes a finalize operation and wherein the plurality of processing functions further include a principal function corresponding to performance of the finalize operation.

### Claim 7 (depends on 6)

The system of claim 6, wherein the plurality of processing elements includes a subset of processing elements that perform the finalize operation and wherein the finalize operation generates a result.

### Claim 8 (depends on 7)

The system of claim 7, wherein the broadcast operation operates to send the result of the finalize operation to the subset of the plurality of processing elements performing corresponding kernel functions.

### Claim 9 (depends on 8)

The system of claim 8, wherein the apply operation applies results of the finalize operation to the subset of the plurality of processing elements performing corresponding kernel functions.

### Claim 10 (depends on 9)

The system of claim 9, wherein the apply operation is performed by a single one of the plurality of processing elements.

### Claim 11 (depends on 9)

The system of claim 9, wherein the apply operation is performed by a selected ones of the plurality of processing elements.

### Claim 12 (depends on 11)

The system of claim 11, wherein the plurality of processing elements includes a plurality of kernel patches, each kernel patch including a plurality of kernels, the plurality of operations further include a complete operation and an echo operation wherein the apply operation is implemented by performance of the complete operation by a single one of the plurality of processing elements from each kernel patch and performance of the echo operation by other ones of the plurality if processing elements.

### Claim 13 (depends on 1)

The system of claim 1, wherein the computational engine is implemented via a wafer-scale engine.

### Claim 14 (depends on 1)

The system of claim 1, wherein each of the plurality of processing elements has a dedicated memory.

### Claim 15 (depends on 14)

The system of claim 14, wherein the computational engine lacks global memory that is shared by the each of the plurality of processing elements.

### Claim 16 (depends on 2)

The system of claim 2, wherein the task input and task output correspond to a quadratic unconstrained binary optimization over a binary vector v=(v1, v2, . . . vn).

### Claim 17 (depends on 16)

The system of claim 16, wherein n of the plurality of processing elements performing corresponding kernel functions each store a current value of vj, where j=1 . . . n.

### Claim 18 (depends on 17)

The system of claim 17, wherein each of the plurality of processing elements performing corresponding kernel functions performs a produce operation that computes a change in cost that would occur if the current value of vj was inverted for j=1 . . . n, wherein the reduce operation generates an update to the current value of vj by selecting a single vj that generated a most favorable cost when inverted, wherein the broadcast operation sends the update to the current value of vj to each of the plurality of processing elements performing corresponding kernel functions and wherein the apply functions applies the update to the current value of vj to each of the plurality of processing elements performing corresponding kernel functions.

### Claim 19

A method comprising:
providing a computational engine having a plurality of processing elements arrayed in accordance with a grid topology, wherein each of the plurality of processing elements is configured to perform one of a plurality of processing functions, and wherein the plurality of processing functions include a principal function, a kernel function, a routing function, and a reducer function;
receiving a task input via a host processor; and
controlling, via the host processor, the computational engine in accordance with an iterative reduce-broadcast-apply process to generate a task output that is based on the task input.

### Claim 20 (depends on 19)

The method of claim 19, wherein the task input and task output correspond to a quadratic unconstrained binary optimization over a binary vector v=(v1, v2, . . . vn).
