NVIDIA Corp v. Neural Ai LLC

1. Case Identification

• Case #: IPR2025-00608
• Patent #: RE49,461
• Filed: March 6, 2025
• Petitioner(s): NVIDIA Corp
• Patent Owner(s): Neural AI, LLC
• Challenged Claims: 21-41

2. Patent Overview

• Title: System and Method for General-Purpose Computation Using Graphics Processors
• Brief Description: The ’461 patent describes systems and methods for performing general-purpose computations, specifically artificial neural networks, on graphics processing units (GPUs). The alleged invention involves managing separate "user interaction" and "computational" execution streams on a CPU and GPU, respectively, and controlling the data exchange between them to prevent data corruption.

3. Grounds for Unpatentability

Ground 1: Obviousness over Buck and Wilt - Claims 21-41 are obvious over Buck in view of Wilt.

• Prior Art Relied Upon: Buck (Application # 2005/0197977) and Wilt (Application # 2003/0140179).
• Core Argument for this Ground:
  • Prior Art Mapping: Petitioner argued that Buck taught a computer system using a GPU to accelerate artificial neural network computations, disclosing a CPU, a GPU (processing unit), and partitioned memory. However, Buck did not focus on task scheduling. Wilt allegedly supplied the missing elements by disclosing an efficient resource management and scheduling system for a host CPU and co-processor GPU. Wilt’s system used a scheduler to construct, queue, and dispatch command buffers (a computational stream) to the GPU, thereby managing multitasking and preventing the CPU from interfering with GPU operations once execution began.
  • Motivation to Combine: A person of ordinary skill in the art (POSITA) would combine the teachings to improve the performance of Buck’s GPU-accelerated neural network system. Implementing Wilt’s established scheduling and task management methods would optimize efficiency and manage data dependencies, a known need for computationally intensive applications. Petitioner asserted this combination was particularly obvious because both Buck and Wilt were assigned to Microsoft, addressed similar CPU/GPU co-processing environments, and were complementary in focus.
  • Expectation of Success: A POSITA would have a reasonable expectation of success, as the combination involved applying a known resource management technique (Wilt) to a known application (Buck's GPU-accelerated system) to achieve the predictable result of efficient, well-managed computation.

Ground 2: Obviousness over Buck, Wilt, and nnet - Claims 23-26, 34-36, and 38-40 are obvious over Buck in view of Wilt and nnet.

• Prior Art Relied Upon: Buck (Application # 2005/0197977), Wilt (Application # 2003/0140179), and nnet (Neural Network Toolbox User’s Guide Version 4).
• Core Argument for this Ground:
  • Prior Art Mapping: This ground built upon the Buck/Wilt combination by adding nnet, a user manual for MATLAB’s Neural Network Toolbox. Petitioner argued that nnet disclosed a standard, user-friendly graphical user interface (GUI) for creating, training, simulating, and editing artificial neural networks. This teaching allegedly mapped directly to dependent claim limitations requiring user-interactive control over setting and editing computational elements, parameters of the network, and parameters of the inputs.
  • Motivation to Combine: A POSITA would be motivated to add a standard GUI, as taught by nnet, to the Buck/Wilt system to provide a user-friendly way to interact with and control the neural network simulations. This was presented as a common and predictable design choice for making complex systems accessible to users.
  • Expectation of Success: Implementing a well-understood feature like a GUI into the known Buck/Wilt system would yield the predictable result of allowing users to easily provide input and control the simulation.

Ground 3: Obviousness over Buck, Wilt, and ANN - Claims 32-37 are obvious over Buck in view of Wilt and ANN.

• Prior Art Relied Upon: Buck (Application # 2005/0197977), Wilt (Application # 2003/0140179), and ANN (Artificial Neural Network Computation on Graphic Process Unit).

• Core Argument for this Ground:

  • Prior Art Mapping: This ground added the teachings of ANN, a publication disclosing the application of GPUs for real-time artificial neural network computation, specifically for object recognition and tracking using real-time video data from a camera. Petitioner argued this mapped to claim limitations requiring a system to include a camera to acquire input data and to generate inputs during the execution of computations. ANN taught that the CPU could receive video data and transfer it to the GPU for processing.
  • Motivation to Combine: A POSITA would be motivated to implement the real-time computer vision application from ANN on the optimized Buck/Wilt GPU-accelerated system. The Buck/Wilt system was specifically designed for the type of parallel computations required by ANN’s algorithm, making it an ideal platform.
  • Expectation of Success: A POSITA would have readily understood how to use the Buck/Wilt system to implement the GPU-based algorithm of ANN and would expect success because the system was configured to handle such computationally intensive, data-parallel tasks.

• Additional Grounds: Petitioner asserted that claims 32-37 were also obvious over Buck, Wilt, and GPU Gems (a book on GPU programming techniques that also teaches real-time computer vision from a camera). Further grounds argued that claims 34-36 were obvious over combinations including nnet with either ANN or GPU Gems, based on similar design modification theories.

4. Arguments Regarding Discretionary Denial

• §314(a) / Fintiv Factors: Petitioner argued against discretionary denial under Fintiv, asserting that the factors strongly favor institution. Key arguments included that the scheduled trial in the parallel district court litigation is more than 18 months after the petition filing and well after the deadline for a Final Written Decision (FWD). Petitioner also stipulated that it would not pursue the same invalidity grounds in the litigation if an inter partes review (IPR) is instituted.
• §325(d) Factors: Petitioner contended that denial under §325(d) is unwarranted because the core prior art references—Buck, Wilt, nnet, and GPU Gems—were never before the examiner during prosecution. While ANN was of record, it was part of a large information disclosure statement and was not substantively applied by the examiner in any rejection. Therefore, Petitioner argued the petition raises new, meritorious arguments that were not previously considered.

5. Relief Requested

• Petitioner requests institution of an IPR and cancellation of claims 21-41 of the ’461 patent as unpatentable under 35 U.S.C. §103.