PTAB

IPR2025-00610

NVIDIA Corp v. LOwensTein Weatherwax LLP

Log In
Key Events
Petition
petition

1. Case Identification

2. Patent Overview

  • Title: System for General-Purpose Computing on Graphics Processing Units
  • Brief Description: The ’438 patent relates to a system for performing general-purpose computing on graphics processing units (GPUs). The challenged reissue claims are directed specifically to using a GPU-based accelerator to perform a sequence of computations representing an artificial neural network, including methods for transferring data between the accelerator and main memory while the GPU is computing.

3. Grounds for Unpatentability

Ground 1: Claims 1-14, 16-34, and 40-54 are obvious over Kirk in view of Oh.

  • Prior Art Relied Upon: Kirk (Patent 7,139,003) and Oh (a 2004 journal article titled "GPU implementation of neural networks").
  • Core Argument for this Ground:
    • Prior Art Mapping: Petitioner argued that Kirk discloses a GPU-accelerated parallel processing system with a host CPU, main memory, and a graphics subsystem (accelerator) containing a programmable GPU and local memory. This system architecture maps to the core components of independent claims 1, 12, 21, 40, and 44. Oh discloses the specific application missing from Kirk: implementing a multi-layer artificial neural network on a commodity GPU to perform real-time text detection in video. Oh’s process, where the output of one neural layer becomes the input for the next, teaches the claimed "sequence of computations representing an artificial neural network" that yields intermediate results.
    • Motivation to Combine (for §103 grounds): A person of ordinary skill in the art (POSITA) would combine Oh’s known neural network technique with Kirk’s advanced GPU architecture to achieve greater computational efficiency. Oh itself sought faster real-time processing, and applying its method to a more powerful, parallel-processing system like Kirk’s was a predictable path to improving performance for a computationally intensive task.
    • Expectation of Success (for §103 grounds): A POSITA would have a reasonable expectation of success because Kirk’s architecture was explicitly designed for parallel data processing, and Oh’s neural network computations are inherently parallelizable tasks.

Ground 2: Claims 1-14, 16-34, and 40-54 are obvious over Kirk in view of Oh and Tamura.

  • Prior Art Relied Upon: Kirk (Patent 7,139,003), Oh, and Tamura (Japanese Patent Appl. No. H04-237388A).
  • Core Argument for this Ground:
    • Prior Art Mapping: This ground adds Tamura to address claim limitations related to parallel data transfer and computation, such as transferring input data to the accelerator while the GPU is already processing previous data (as in claims 1[d][iii], 12[b]). Tamura explicitly teaches a neuroprocessor that inputs and writes new data to memory while calculations are being performed on previously inputted data, thereby reducing total processing time.
    • Motivation to Combine (for §103 grounds): A POSITA seeking to optimize the real-time performance of the Kirk/Oh system would be motivated to incorporate Tamura’s efficiency-improving technique. To avoid processing bottlenecks and keep up with a real-time data stream (like video in Oh), it would be an obvious step to overlap data transfers with ongoing computations, a technique Tamura taught was applicable to any neural network processor.
    • Expectation of Success (for §103 grounds): A POSITA would expect this combination to succeed in improving efficiency, as Tamura’s method was a well-understood solution to the known problem of sequential processing delays in neural networks.

Ground 3: Claims 40-43 are obvious over Kirk in view of Oh and GPU Gems.

  • Prior Art Relied Upon: Kirk (Patent 7,139,003), Oh, and GPU Gems 2 (a 2005 publication on GPU programming techniques).

  • Core Argument for this Ground:

    • Prior Art Mapping: This ground adds GPU Gems to address limitations in claims 40-43, which recite using pointers to manage data in memory partitions and swapping pointers so that an output becomes an input for a subsequent computation layer. GPU Gems is a textbook that describes state-of-the-art GPU programming, including the common technique of "pointer swapping" (also known as "double buffering") to efficiently handle iterative computations without physically copying large amounts of data.
    • Motivation to Combine (for §103 grounds): A POSITA implementing Oh's multi-layer neural network on Kirk's hardware would need an efficient way to pass the output of one layer to the next. Pointer swapping, as detailed in GPU Gems, was a fundamental and well-known technique in GPU programming for precisely this purpose. A POSITA would have been motivated to use this standard, efficient method rather than a more cumbersome alternative.
    • Expectation of Success (for §103 grounds): Success was expected because pointer swapping was a routine and widely implemented technique in GPGPU applications, and applying it to manage data flow in a neural network was a straightforward implementation of a known programming paradigm.

  • Additional Grounds: Petitioner asserted an additional obviousness challenge against claims 40-43 based on the combination of Kirk, Oh, Tamura, and GPU Gems, which relied on the same combination of rationales presented in the grounds above.

4. Arguments Regarding Discretionary Denial

  • Petitioner argued discretionary denial would be inappropriate. The petition included a stipulation that, if the inter partes review (IPR) is instituted, Petitioner will not pursue in the co-pending district court litigation the same grounds or any grounds that could have reasonably been raised in the petition, addressing potential concerns under Fintiv.

5. Relief Requested

  • Petitioner requests institution of an IPR and cancellation of claims 1-14, 16-34, and 40-54 of Patent RE48,438 as unpatentable under 35 U.S.C. §103.