# US 2012/0254604 A1 · Run-Ahead Approximated Computations

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

- Number: US 2012/0254604 A1
- Document ID: US20120254604A1
- Status: Published application
- Kind: A1
- Application number: 13/468,064
- Filed: 2012-05-10
- Published: 2012-10-04
- Earliest claimed: 2011-03-29
- Assignees: International Business Machines Corporation
- Inventors: Christopher J. Craik; H. Peter Hofstee; Damir A. Jamsek; Jian Li
- Web page: https://patentgrep.com/patent/US20120254604A1

## CPC

- 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: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Mechanisms are provided for performing approximate run-ahead computations. A first group of compute engines is selected to execute full computations on a full set of input data. A second group of compute engines is selected to execute computations on a sampled subset of the input data. A third group of compute engines is selected to compute a difference in computation results between first computation results generated by the first group of compute engines and second computation results generated by the second group of compute engines. The second group of compute engines is reconfigured based on the difference generated by the third group of compute engines.

## Claims

### Claim 1

A method, in a data processing system, for performing approximate run-ahead computations, comprising:
selecting a first group of compute engines, from a pool of compute engines, to execute full computations on a full set of input data;
selecting a second group of compute engines, from the pool of compute engines, to execute computations on a sampled subset of the input data;
selecting a third group of compute engines, from the pool of compute engines, to compute a difference in computation results between first computation results generated by the first group of compute engines and second computation results generated by the second group of compute engines; and
reconfiguring the second group of compute engines based on the difference generated by the third group of compute engines.

### Claim 2 (depends on 1)

The method of claim 1, wherein the pool of compute engines comprise at least one of a pool of processors, a pool of functional units within one or more processors, a pool of data processing devices, each data processing device comprising one or more processors, or a pool of special purpose accelerators that accelerate a particular type of computation.

### Claim 3 (depends on 1)

The method of claim 1, wherein the second group of compute engines comprises a number of compute engines that is smaller in number than a number of compute engines in the first group of compute engines.

### Claim 4 (depends on 1)

The method of claim 1, wherein the second group of compute engines execute approximate computations on the sampled subset of the input data.

### Claim 5 (depends on 4)

The method of claim 4, wherein reconfiguring the second group of compute engines based on the difference generated by the third group of compute engines comprises reconfiguring the second group of compute engines based on a degree of accuracy of the approximate computations and a measure of confidence in these approximate computations.

### Claim 6 (depends on 1)

The method of claim 1, wherein selecting a first group of compute engines, from a pool of compute engines, to execute full computations on a full set of input data comprises selecting a group of compute engines from the pool of compute engines based on at least one of an arbitrary selection method, a selection of compute engines based on performance capabilities of the compute engines, current workloads of the compute engines, or physical affinity of the compute engines.

### Claim 7 (depends on 1)

The method of claim 1, wherein the second group of compute engines execute the computations in a runahead manner such that the data upon which the second group of compute engines operates comprises data that is relatively later in sequence in the set of input data than data upon which the first group of compute engines operates.

### Claim 8 (depends on 1)

The method of claim 1, wherein reconfiguring the second group of compute engines comprises increasing a number of compute engines in the second group of compute engines.

### Claim 9 (depends on 1)

The method of claim 1, wherein reconfiguring the second group of compute engines further comprises increasing a size of the sampled subset of data from the set of input data upon which the second group of compute engines operates.

### Claim 10 (depends on 1)

The method of claim 1, wherein the reconfiguring of the second group of compute engines is performed in response to the difference in computation results not being within a predetermined tolerance specified by one or more threshold values.

### Claim 11 (depends on 10)

The method of claim 10, wherein in response to the difference in computation results not being within the predetermined tolerance, the method further comprises initiating a partial re-computation by the reconfigured second group of compute engines in a next phase of computation.

### Claim 12 (depends on 1)

The method of claim 1, wherein, in response to the difference in computation results being within a predetermined tolerance as specified by one or more threshold values, the method further comprises:
generating an interrupt to interrupt the computations being performed by the first group of compute engines; and
utilizing results of the computations of the second group of compute engines as results of a computation phase of code being executed by the first and second groups of compute engines.

### Claim at position 13

13-25. (canceled)
