Singular Sues Google for $1.67B: TPU Patent Clash

On the 12th of January, news emerged that Singular Computing's computer scientist, Joseph Bates, has filed a lawsuit against Google, alleging the infringement of two technology patents associated with Google's Tensor Processing Unit (TPU). The claim is seeking substantial damages amounting to $1.67 billion.

Bates' patents, initially applied for in 2009 and publicly disclosed in 2010, delineate a computer architecture designed to execute a significant volume of low-precision calculations in each processor cycle. While such lower precision may seem impractical for traditional computing workloads, the lawsuit contends that it is well-suited for artificial intelligence software capable of adapting to this lower precision.

Singular Computing underscores that these patented technologies are not merely theoretical. Prototypes based on these designs were reportedly developed shortly after the first patent application. The relevant patents include the initial patent number US 8,407,273, along with subsequent patent numbers US 9,218,156 and US 10,416,961.

Bates' legal representative, Kerry Timbers, alleges that Google intentionally incorporated Bates' architecture into its TPU v2 and v3 processors without permission, deliberately infringing on the associated patents. Initially seeking $7 billion in damages, Singular Computing adjusted this claim to $1.67 billion during the trial.

In response, Google's legal team argues that those involved in the development of Google TPU had no connection to Bates or his blueprints. Google repeatedly denies the patent infringement allegations, with public relations pointing to another ongoing case in the U.S. appeals court this week, where Google is arguing why Singular Computing's patents should be considered invalid. Essentially, Google is seeking to have the patents declared invalid to dismiss Singular's infringement lawsuit.

Google introduced the TPU in 2016 to support machine learning features in products like Gmail, Google Maps, and YouTube. At a higher level, the accelerators are essentially a stack of brain-inspired matrix math engines called MXUs, supported by high-bandwidth memory and CPU cores for programmability.

Currently, Google is unveiling the fifth generation of TPU chips, positioning them as alternatives to AI GPUs for cloud-based artificial intelligence training and inference workloads. The trial is set to commence next Monday and is expected to last for at least two weeks.