Google and UC Berkeley TPU Paper: Five Generations of Tensor Processing Units (v2 to Ironwood) in IEEE Micro – News and Statistics

Here are five key points from the technical paper on Google’s Tensor Processing Units (TPUs):

1. The paper provides a comprehensive overview of five generations of Google’s TPUs, from the TPU v2 to the latest Ironwood system. It details how the TPU platform evolved to become more scalable, resilient, power-efficient, and sustainable.

2. The authors report significant improvements in high-bandwidth memory capacity, bandwidth per node, peak node performance, and overall supercomputer performance over an eight-year period.

3. The paper highlights several technical innovations that contributed to system resilience, including the use of optical circuit switches, built-in self-test mechanisms, and hardware replay capabilities.

4. Researchers note advancements in performance per watt and reductions in carbon emissions per floating-point operation, making TPUs more efficient and sustainable platforms for artificial intelligence training.

5. The study concludes by identifying six features that are likely to characterize successful training accelerators in the coming decade.

The full study is available in the July/August 2026 issue of IEEE Micro.