NVIDIA Corp v. Neural Ai LLC

1. Case Identification

• Case #: IPR2025-00606
• Patent #: 8,648,867
• Filed: February 24, 2025
• Petitioner(s): NVIDIA Corporation
• Patent Owner(s): Neural AI, LLC
• Challenged Claims: 1-19

2. Patent Overview

• Title: System for Numerical Simulation
• Brief Description: The ’867 patent discloses a system for performing general-purpose computing, such as numerical simulations, on graphics processing units (GPUs). The purported invention is an architecture that uses a dedicated "accelerator controller" to manage GPU computations, thereby freeing the main central processing unit (CPU) for other tasks.

3. Grounds for Unpatentability

Ground 1: Claims 1-19 are obvious over Nickolls in view of GPU Gems

• Prior Art Relied Upon: Nickolls (Patent 7,861,060) and GPU Gems (a 2005 book titled GPU Gems 2: Programming Techniques for High-Performance Graphics and General-Purpose Computation).
• Core Argument for this Ground:
  • Prior Art Mapping: Petitioner argued that Nickolls, an NVIDIA patent, discloses a GPU architecture for parallel data processing that meets the core limitations of the challenged claims. Specifically, Nickolls’s "core interface 128" and "instruction unit 312" allegedly perform the functions of the claimed "accelerator controller" by managing the setup and control of computations on the GPU’s processing cores, separate from the host CPU. Nickolls further taught using the output of one computational step ("intermediate results") as the input for a subsequent step. While Nickolls disclosed the necessary hardware architecture, it did not explicitly teach the memory management technique of swapping pointers. Petitioner asserted that GPU Gems, a well-known NVIDIA programming guide, supplied this element. GPU Gems explicitly described and provided code for using a pointer-swapping technique (also known as "double buffering" or "ping-pong rendering") in a numerical simulation on a GPU. This technique involved swapping pointers to input and output memory buffers (currentTarget and currentSource) at the end of each computational cycle, directly teaching the limitation of "swapping the first pointer with the second pointer" as recited in independent claim 16.
  • Motivation to Combine: Petitioner contended a person of ordinary skill in the art (POSITA) would combine these references. Nickolls provided an advanced GPU architecture designed for general-purpose computations like "modeling and simulation." GPU Gems provided established programming techniques to efficiently implement such simulations on GPUs. A POSITA would combine the well-known pointer-swapping technique from GPU Gems with the hardware of Nickolls to achieve the predictable result of efficient, high-performance numerical simulation by avoiding the costly process of physically moving large data arrays in memory between computational cycles.
  • Expectation of Success: A POSITA would have had a high expectation of success. The Nickolls system was designed for parallel processing with a shared memory architecture suitable for iterative calculations, and pointer swapping was a fundamental and reliable programming technique widely used to manage memory efficiently.

Ground 2: Claims 1-19 are obvious over Kirk in view of GPU Gems

• Prior Art Relied Upon: Kirk (Patent 7,139,003) and GPU Gems.
• Core Argument for this Ground:
  • Prior Art Mapping: Petitioner presented a similar argument using Kirk, another NVIDIA patent, as the primary reference. Kirk allegedly disclosed a "Computing System" with a programmable graphics processor for general data processing. Petitioner identified Kirk’s "Graphics Interface 117," "Front End 1230," and "Index Processor 1235" as collectively constituting the claimed "accelerator controller," as they receive and format commands from the host processor for execution by the GPU pipeline. Kirk also disclosed that data produced in one processing pass could be written to a buffer to be read during a "subsequent pass." As with Nickolls, Kirk did not explicitly teach pointer swapping to implement this iterative process. GPU Gems was again relied upon to teach this well-known technique, providing the code and explanation for swapping pointers between input and output buffers to make the output of one cycle the input for the next.
  • Motivation to Combine: The motivation to combine Kirk and GPU Gems was analogous to that for Nickolls. Kirk disclosed a high-performance GPU, and GPU Gems was a contemporaneous guide on programming such hardware. A POSITA seeking to implement efficient numerical simulations on the Kirk hardware would naturally turn to the techniques described in GPU Gems. The combination would have been a straightforward application of a known software technique (pointer swapping) to a suitable hardware system (Kirk's GPU) to achieve the predictable benefits of improved performance and efficient memory management.
  • Expectation of Success: Success was expected because Kirk’s architecture supported multi-pass operations and had the necessary memory structures for implementing the double-buffering technique taught by GPU Gems. The combination involved applying a routine programming optimization to a system designed for such parallel computations.

4. Arguments Regarding Discretionary Denial

• Petitioner argued that discretionary denial under 35 U.S.C. §314(a) is not warranted. It asserted that it would stipulate not to pursue in the parallel district court litigation the same grounds raised in the petition, or any grounds that could have reasonably been raised, consistent with the PTAB's guidance in Sotera Wireless.
• Petitioner also argued that denial under 35 U.S.C. §325(d) is unwarranted because none of the prior art references relied upon in the petition (Nickolls, Kirk, and GPU Gems) were considered by the USPTO during the original prosecution of the ’867 patent.

5. Relief Requested

• Petitioner requested the institution of an inter partes review and the cancellation of claims 1-19 of Patent 8,648,867 as unpatentable under 35 U.S.C. §103.