# US 2021/0342286 A1 · Cgra accelerator for weather/climate dynamics simulation

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

- Number: US 2021/0342286 A1
- Document ID: US20210342286A1
- Status: Published application
- Kind: A1
- Application number: 16/866,121
- Filed: 2020-05-04
- Published: 2021-11-04
- Earliest claimed: 2020-05-04
- Assignees: International Business Machines Corporation; ETH Zurich
- Applicants: International Business Machines Corporation; ETH Zurich
- Inventors: Ronald Peter Luijten; Gagandeep Singh; Joost VandeVondele
- Web page: https://patentgrep.com/patent/US20210342286A1

## CPC

- G06F 13/4027: {using bus bridges (G06F13/4022 takes precedence)}
- G06F 17/13: Differential equations (using digital differential analysers G06F7/64)
- Y02A 90/10: Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- Y02A: TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02D: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE

## Abstract

A coarse-grained reconfigurable array accelerator for solving partial differential equations for problems on a regular grid is provided. The regular grid comprises grid cells which are representative for a physical natural environment wherein a list of physical values is associated with each grid cell. The accelerator comprises configurable processing elements in an accelerator-internal grid connected by an accelerator-internal interconnect system and memory arrays comprising memory cells. The memory arrays are connected to the accelerator-internal interconnect system. Selected ones of the memory arrays are positioned within the accelerator corresponding to positions of the grid cells in the physical natural environment. Thereby, each group of the memory cells is adapted for storing the list of physical values of the corresponding grid cell of the physical natural environment.

## Claims

### Claim 1

A coarse-grained reconfigurable array accelerator for solving partial differential equations for problems on a regular grid, said regular grid comprising grid cells being representative for a physical natural environment, wherein a list of physical values is associated with each grid cell, wherein said accelerator is adapted for solving said partial differential equations using a plurality of stencil computing kernels, wherein a horizontal compute stencil kernel has operand access in an x-y plane, and wherein a vertical compute stencil kernel has operand access in a plane vertical to said x-y plane, and wherein a determination of horizontal stencil computation results and vertical stencil computation results alternates, said accelerator comprising
configurable processing elements in an accelerator-internal grid connected by an accelerator-internal interconnect system,
memory arrays comprising memory cells, said memory arrays being connected to said accelerator-internal interconnect system, wherein selected ones of said memory arrays are positioned within said accelerator corresponding to positions of said grid cells in said physical natural environment, and wherein each group of said memory cells is adapted for storing said list of physical values of said corresponding grid cell of said physical natural environment.

### Claim 2 (depends on 1)

The accelerator according to claim 1, wherein said positioning of said memory arrays is equivalent to an x-y positioning of said grid cells in said physical natural environment.

### Claim 3 (depends on 1)

The accelerator according to claim 1, wherein said list of physical values of a selected grid cell of said physical natural environment is mapped to multiple memory arrays.

### Claim 4 (depends on 1)

The accelerator according to claim 1, wherein said memory arrays have multiple read and write ports.

### Claim 5 (depends on 1)

The accelerator according to claim 1, wherein said accelerator comprises a first external bus adapted for connecting an external accelerator memory.

### Claim 6 (depends on 1)

The method according to claim 1, wherein said accelerator comprises a second external bus adapted for connecting a host computer.

### Claim 7

(canceled)

### Claim 8

(canceled)

### Claim 9 (depends on 1)

The accelerator according to claim 1, wherein said accelerator is adapted to be reconfigured between different stencil compute kernels.

### Claim 10 (depends on 1)

The accelerator according to claim 1, wherein said accelerator is adapted for an overlapping of data fetching from an accelerator external memory for a next stencil computing cycle and a current stencil computing cycle across said entire grid.

### Claim 11

A method for solving partial differential equations using a coarse-grained reconfigurable array accelerator for problems on a regular grid, said regular grid comprising grid cells being representative for a physical natural environment, wherein a list of physical values is associated with each grid cell, wherein a horizontal compute stencil kernel has operand access in an x-y plane, and wherein a vertical compute stencil kernel has operand access in a plane vertical to said x-y plane,
wherein said configurable processing elements are connected in an accelerator-internal grid by an accelerator-internal interconnect system,
wherein memory arrays comprising memory cells, said memory arrays being connected to said accelerator-internal interconnect system, said method comprising
positioning selected ones of said memory arrays said accelerator such that they correspond to positions of said grid cells in said physical natural environment,
storing said list of physical values of said corresponding grid cell of said physical natural environment in each group of said memory cells,
determining horizontal stencil computation results and vertical stencil computation results alternatively, and
solving said PDEs using a plurality of stencil computing kernels using said accelerator.

### Claim 12 (depends on 11)

The method according to claim 11, wherein said positioning of said memory arrays is equivalent to an x-y positioning of said grid cells in said physical natural environment.

### Claim 13 (depends on 11)

The method according to claim 11, also comprising
mapping said list of physical values of a selected grid cell of said physical natural environment to multiple memory arrays.

### Claim 14 (depends on 11)

The method according to claim 11, wherein said memory arrays have multiple read and write ports.

### Claim 15 (depends on 11)

The method according to claim 11, also comprising
connecting a first external bus of said accelerator to an external accelerator memory.

### Claim 16 (depends on 11)

The method according to claim 11, also comprising
connecting a second external bus of said accelerator to a host computer.

### Claim 17

(canceled)

### Claim 18

(canceled)

### Claim 19 (depends on 11)

The method according to claim 11, also comprising
reconfiguring said accelerator different stencil compute kernels.

### Claim 20

A computer program product for solving partial differential equations using a coarse-grained reconfigurable array accelerator for problems on a regular grid, said regular grid comprising grid cells being representative for a physical natural environment, wherein a list of physical values is associated with each grid cell, wherein a horizontal compute stencil kernel has operand access in an x-y plane, and wherein a vertical compute stencil kernel has operand access in a plane vertical to said x-y plane, wherein said configurable processing elements are connected to an accelerator-internal grid by an accelerator-internal interconnect system, and wherein memory arrays comprising memory cells, said memory arrays being connected to said accelerator-internal interconnect system,
the computer program product comprising a computer readable storage medium having program instructions embodied therewith, said program instructions being executable by one or more computing systems or controllers to cause said one or more computing systems to
position selected ones of said memory arrays said accelerator such that they correspond to positions of said grid cells in said physical natural environment,
store said list of physical values of said corresponding grid cell of said physical natural environment in each group of said memory cells,
determining horizontal stencil computation results and vertical stencil computation results alternatively, and
solving said PDEs using a plurality of stencil computing kernels using said accelerator.
