# US 2025/0095100 A1 · Methods and systems for expanding gpu memory footprint based on hybrid-memory

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

- Number: US 2025/0095100 A1
- Document ID: US20250095100A1
- Status: Published application
- Kind: A1
- Application number: 18/764,154
- Filed: 2024-07-04
- Published: 2025-03-20
- Earliest claimed: 2021-12-27
- Inventors: VENKATACHARY SRINIVASAN; DAVID MICHAEL FINNEGAN
- Web page: https://patentgrep.com/patent/US20250095100A1

## CPC

- G06T 1/60: Memory management
- G06F 16/27: Replication, distribution or synchronisation of data between databases or within a distributed database system; Distributed database system architectures therefor
- G06T 1/20: Processor architectures; Processor configuration, e.g. pipelining
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

In one aspect, a computerized method for expanding a graphics processing unit (GPU) memory footprint based on a hybrid-memory of a distributed database system (DDBS) includes the step of providing the DDBS. The DDBS is modified to include a plurality of GPUs; providing a local memory of a GPU of the plurality of GPUs. The method includes the step of filling the local memory of the GPU with one or more digests from the DDBS. The method includes the step of running a distributed general-purpose cluster-computing framework instance on the local memory of the GPU. The method includes the step of fetching data from the local memory of the GPU using the distributed general-purpose cluster-computing framework instance. The method includes the step of storing a result of the fetch operation in the DDBS to extend the local memory of the GPU to handle more data than what is fitted into the local memory of the GPU.

## Claims

### Claim 1

A computerized method for expanding a graphics processing unit (GPU) memory footprint based on a hybrid-memory of a distributed database system (DDBS) comprising:
providing the DDBS, wherein the DDBS is modified to include a plurality of GPUS;
providing a local memory of a GPU of the plurality of GPUs;
filling the local memory of the GPU with one or more digests from the DDBS;
running a distributed general-purpose cluster-computing framework instance on the local memory of the GPU;
fetching data from the local memory of the GPU using the distributed general-purpose cluster-computing framework instance; and
storing a result of the fetch operation in the DDBS to extend the local memory of the GPU to handle more data than what is fitted into the local memory of the GPU.

### Claim 2 (depends on 1)

The computerized method of claim 1, wherein the DDBS comprises a No-SQL DDBS.

### Claim 3 (depends on 1)

The computerized method of claim 1, wherein a plurality of GPUs are provided.

### Claim 4 (depends on 3)

The computerized method of claim 3, wherein a distributed general-purpose cluster-computing framework process is run on each of the plurality of GPUs.

### Claim 5 (depends on 4)

The computerized method of claim 4, wherein the distributed general-purpose cluster-computing framework comprises an open-source distributed general-purpose cluster-computing framework.

### Claim 6 (depends on 5)

The computerized method of claim 5, wherein the open-source distributed general-purpose cluster-computing framework comprises an APACHE SPARK distributed general-purpose cluster-computing framework.

### Claim 7 (depends on 6)

The computerized method of claim 6, wherein DDBS uses a hybrid memory architecture that provides an ability to have real-time access to data by leveraging one or more flash memory systems.

### Claim 8 (depends on 7)

The computerized method of claim 7, wherein the DDBS comprises a hybrid memory architecture that is extended to the plurality of GPUs.

### Claim 9 (depends on 8)

The computerized method of claim 8 further comprising:
using a two-faced process that distributes the data in parallel to the local memory of each GPU of the plurality of GPUs.

### Claim 10 (depends on 9)

The computerized method of claim 9 further comprising:
with the DDBS, running a divide and conquer algorithm to run a computation on each of GPU of the plurality of GPUs to analyze the data that each GPU has in its respective local memory.

### Claim 11 (depends on 10)

The computerized method of claim 10, wherein the DDBS runs a Spark instance on each local memory of each GPU of the plurality of GPS.

### Claim 12 (depends on 11)

The computerized method of claim 11, wherein the DDBS store digests of a larger piece of data in a DDBS node memory.

### Claim 13 (depends on 12)

The computerized method of claim 12, wherein the DDBS implements a one-way hash RIPEMD 160 to produce a digest that is stored as a set of relevant data components for processing by the DDBS.

### Claim 14 (depends on 13)

The computerized method of claim 13 further comprising scanning the DDBS and obtain the digest; and
using the digest obtained from the DDBS to populate each local memory of each GPU of the plurality of the GPUs.

### Claim 15 (depends on 14)

The computerized method of claim 14, wherein when running an individual process on a specified GPU of the plurality of GPUs, using the digest, the data from the DDBS 100 is fetched in batches.