Licensing TPU IP Cores: The Case for On-Premises Solutions

Leverage Proven AI and ML Architecture

Chiplets are modular silicon pieces designed for tasks like processing, memory, or I/O. Unlike traditional single-die chips, they allow designers to combine pre-built modules. Advanced packaging technologies, like 2.5D interposers or 3D stacking, connect these chiplets into high-performance systems.

TPUs are purpose-built for AI and ML tasks, optimized for tensor operations. Tensors are mathematical objects that generalize scalars, vectors, and matrices to higher dimensions, making them essential for representing and manipulating the large datasets used in AI and ML. Operations on tensors, such as matrix multiplications, form the backbone of neural network computations, central to training and inference. Licensing a TPU IP core allows fabless ASIC companies to integrate a proven architecture into their own custom hardware, avoiding the time and costs of designing an AI accelerator from scratch.

Reduced Time-to-Market

Designing a custom AI accelerator is expensive, time-consuming, and engineering resource-intensive. By acquiring TPU IP cores, fabless companies can significantly reduce their time-to-market. This enables them to dedicate their capital and engineering resources to quickly address customer demands and seize market opportunities.

Lower Long-Term Costs

While the upfront costs of licensing TPU IP and manufacturing ASICs can be high, the long-term expenses of operating on-premises TPU solutions can be lower, particularly for companies with high, consistent workloads.

Focus on Differentiation

By licensing verified core TPU IP customized for AI and ML accelerator designs, fabless ASIC companies can focus their resources on areas that differentiate their products, such as system integration, software support, or optimizing other on-chip components.

Notable TPU IP Vendors

Often, companies the develop and design TPUs consider them to be the crown-jewel of their technology portfolio, and will not license their IP. For example, as of this writing, Google offers their TPU technology only as a cloud-based service. However, a handful of vendors offer TPU IP cores. As of this writing, Arm and Tachyum provide licensable solutions. These cores are designed for efficient AI and ML acceleration, offering scalability and adaptability for various workloads.

Making the Right Choice

The decision to license TPU IP cores depends on the company’s specific needs:

• Workload Consistency: For constant, high-volume workloads, owning TPU-based hardware can reduce operational expenses over time.

• Budget Constraints: Companies with limited upfront capital can strategically license IP to balance costs and development speed.

• Customization Needs: Companies requiring tailored hardware solutions for specific AI and ML workloads benefit from the flexibility of licensed TPU IP cores.

Copyright 2025. All rights reserved.

Saman Taherian

Lapin & Taherian

Attorneys At Law