Google Unveils TPU 3.0 With Liquid Cooling and Modest Gains
Google reveals its new TPU 3.0 AI coprocessor, which requires liquid cooling and promises up to eight times the performance of its predecessor. Despite the new naming convention, the architectural changes appear to be relatively minor.
Google reveals its next-generation TPU 3.0 AI coprocessor at its annual IO conference, highlighting a shift from version notation to revision notation. Despite the updated naming convention, a closer look at the architecture suggests the step from TPUv2 to TPUv3 is relatively minor and perhaps more accurately resembles a stepping stone revision rather than a complete overhaul.
The new TPU 3.0 chip runs so hot that Google implements liquid cooling in its datacenters for the first time. The company claims each TPU 3.0 pod delivers eight times the performance of a TPUv2 pod, pushing past the 100 petaflops mark compared to the previous 11.5 petaflops, though the exact math behind this performance leap raises some questions.
Access to these AI chips remains strictly locked behind Google's TensorFlow deep learning framework, with no direct hardware access granted to external developers. Because Google invests heavily in software optimization for its current beta Cloud TPU deployment, the company heavily retains the existing hardware abstraction layers and APIs to ensure a smooth transition to the new hardware.