PTAB

IPR2025-00609

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 Processors
  • Brief Description: The ’438 patent, a reissue of Patent 9,189,828, describes a computer system using a graphics processing unit (GPU) as an accelerator for general-purpose computing. The system is directed to performing a sequence of computations representing an artificial neural network, featuring a specialized controller to manage data transfers and computations, thereby freeing the central processing unit (CPU) for other tasks.

3. Grounds for Unpatentability

Ground 1: Claims 1-14, 16-34, and 40-54 are obvious over Nickolls in view of ANN

  • Prior Art Relied Upon: Nickolls (Patent 7,861,060) and ANN (Z. Luo, Artificial Neural Network Computation on Graphic Process Unit, 2005).
  • Core Argument for this Ground:
    • Prior Art Mapping: Petitioner argued that the combination of Nickolls and ANN teaches all limitations of the challenged claims. Nickolls disclosed a computer system with a flexible, general-purpose GPU architecture, including a CPU, main memory, and a GPU accelerator with its own memory and controller (a “core interface”). This architecture was explicitly designed for parallel data processing and could be leveraged for general-purpose computations like matrix algebra. ANN taught the specific application claimed in the ’438 patent: implementing a multi-layer artificial neural network on a commodity NVIDIA GPU to perform a real-time task (tracking a soccer ball). ANN disclosed that using a GPU was significantly faster than a CPU for this purpose and taught using intermediate computation results as inputs for subsequent computational layers.
    • Motivation to Combine: A person of ordinary skill in the art (POSITA) seeking to efficiently execute a computationally intensive artificial neural network, as taught by ANN, would combine it with an advanced, state-of-the-art parallel processing system like that in Nickolls. Petitioner asserted this was a straightforward application of a known technique (ANN’s neural network) to a known system (Nickolls’ GPU architecture) to achieve the predictable result of improved performance. The fact that Nickolls was assigned to NVIDIA, a leader in GPUs, and that ANN was implemented on an NVIDIA GPU would have further motivated the combination.
    • Expectation of Success: A POSITA would have had a reasonable expectation of success. Nickolls was designed for parallel computations like matrix algebra, a key operation in the neural networks described by ANN. Implementing the known software methods of ANN on the suitable hardware architecture of Nickolls would predictably result in an efficient system.

Ground 2: Claims 1-14, 16-34, and 40-54 are obvious over Nickolls and ANN, further in view of Tamura and/or GPU Gems

  • Prior Art Relied Upon: Nickolls (Patent 7,861,060), ANN (Z. Luo, 2005), Tamura (JPH04-237388A), and GPU Gems (GPU Gems 2: Programming Techniques for High-Performance Graphics and General-Purpose Computation, 2005).
  • Core Argument for this Ground:
    • Prior Art Mapping: This ground built upon the Nickolls/ANN combination by adding known optimization techniques for data handling that Petitioner argued were recited in various dependent claims. Tamura taught a method to reduce processing time in neural network processors by inputting and writing new data to memory in parallel with ongoing calculations on previously inputted data. This directly addresses the ’438 patent’s limitations regarding overlapping data transfers and computations. GPU Gems, a well-known programming guide, disclosed standard techniques for efficient data management on GPUs, such as "double buffering" and pointer swapping. This technique allows the output of one computation to be used as the input for the next by simply swapping memory pointers, rather than copying large blocks of data, which maps to claim limitations concerning memory partitions and using output data as input for a subsequent layer of neurons.
    • Motivation to Combine: A POSITA, having combined Nickolls and ANN to create a real-time neural network application, would be motivated to further optimize its performance. To handle continuous, real-time data input without stalling the GPU, a POSITA would look to known techniques for parallelizing data transfers and computations, as taught by Tamura. To efficiently manage data flow between computational layers of the neural network, a POSITA would naturally turn to well-established, fundamental GPGPU techniques like the pointer swapping described in GPU Gems to avoid data-copying overhead.
    • Expectation of Success: The techniques in Tamura and GPU Gems were established solutions to common problems in parallel and real-time processing. Applying these known optimization methods to the Nickolls/ANN system would predictably improve its efficiency and throughput with a high expectation of success.

4. Arguments Regarding Discretionary Denial

  • Petitioner stipulated that if inter partes review (IPR) is instituted, it will not pursue in the related district court litigation the same grounds or any grounds that could have reasonably been raised in the petition.

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.