# US 2026/0064599 A1 · Memory allocation in a reconfigurable data processor

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

- Number: US 2026/0064599 A1
- Document ID: US20260064599A1
- Status: Published application
- Kind: A1
- Application number: 19/382,321
- Filed: 2025-11-07
- Published: 2026-03-05
- Earliest claimed: 2023-08-09
- Assignees: SambaNova Systems, Inc.
- Applicants: SambaNova Systems, Inc.
- Inventors: Matthew FELDMAN; Yaqi ZHANG
- Web page: https://patentgrep.com/patent/US20260064599A1

## CPC

- G06F 12/10: Address translation
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/063: using electronic means
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A system comprises an array of reconfigurable units including a plurality of pattern compute units (PCUs) and a plurality of pattern memory units (PMUs), and a compiler configured to receive a tensor including a plurality of memory access patterns of a first type (read) and a second type (write) located in a logical memory of the compiler. Each memory access pattern includes a memory access particular to that type. The compiler is configured to create a plurality of duplicates of the tensor and assign one or more contexts of any type to the duplicate tensors, such that no two contexts of the same type are in the same duplicate tensor.

## Claims

### Claim 1

A method for a data processing system comprising: an array of reconfigurable units including a plurality of pattern compute units (PCUs) and a plurality of pattern memory units (PMUs), each of the plurality of the PMUs further comprising a memory having one or more first type of (read) I/O ports and one or more second type of (write) I/O ports, the method comprising:
receiving by a compiler, a tensor including ‘n’ read memory access patterns or ‘m’ write memory access patterns located in a logical memory of the compiler, a read memory access pattern (read context) including a read memory access (read data region) and a write memory access pattern (write context) including a write memory access (write data region);
creating by the compiler, a plurality of copies of the tensor to create a plurality of duplicate tensors; and
assigning by the compiler, up to ‘n’ read contexts or up to ‘m’ write contexts to one or more duplicate tensors, such that no duplicate tensor has more contexts of a particular type than ports of that type.

### Claim 2 (depends on 1)

The method of claim 1, wherein the first type of context is a read context including read data corresponding to a read operation to be performed in a physical memory location in the array of reconfigurable units.

### Claim 3 (depends on 1)

The method of claim 1, wherein the second type of context is a write context including write data corresponding to a write operation to be performed in a physical memory location in the array of reconfigurable units.

### Claim 4 (depends on 2)

The method of claim 2, wherein the read data includes a first read data portion and a second read data portion and the write data includes a first write data portion and a second write data portion.

### Claim 5 (depends on 4)

The method of claim 4, further comprising: assigning the first read data portion to a first PMU and the second read data portion to a second PMU.

### Claim 6 (depends on 5)

The method of claim 5, further comprising: assigning the first read data portion from the first PMU and the second read data portion from the second PMU to a reorder buffer in any order and placing by the reorder buffer, the first read data portion and the second read data portion in a correct order to generate an ordered read data.

### Claim 7 (depends on 6)

The method of claim 6, further comprising: to reading from a physical memory location in the array of reconfigurable units, corresponding to the ordered read data.

### Claim 8 (depends on 2)

The method of claim 2, further comprising writing to a physical memory location in the array of reconfigurable units, corresponding to the write data.

### Claim 9 (depends on 2)

The method of claim 2, wherein the two or more read contexts are overlapping.

### Claim 10 (depends on 2)

The method of claim 2, wherein the two or more write contexts are overlapping.

### Claim 11 (depends on 1)

The method of claim 1, further including:
trimming by the compiler, a duplicate tensor to retain portions including its corresponding assigned contexts and remove portions that are inconsequential to the assigned contexts; and
dispatching by the compiler the assigned contexts in the duplicate tensor to one or more of the plurality of the PMUs.
