The Jeuns

New Processing Unit Unveiled for Advancing AI Models

May 17, 2024 | by The Jeuns

Google has revealed its latest addition to the tensor processing unit (TPU) lineup at the recent I/O event – Trillium. This new processor is specifically designed to power next-generation AI models, following in the footsteps of its predecessors which were originally developed for internal Google products such as Gmail, Google Maps, and YouTube.

Trillium represents the sixth generation of Google’s TPUs and boasts a significant performance upgrade. With a 4.7x increase in peak compute performance and double the high bandwidth memory capacity compared to the TPU v5e design, this new processor is set to deliver impressive results. According to Google, Trillium is capable of achieving 926 teraFLOPS at BF16 and 1,847 teraFLOPS at INT8, making it twice as fast as the previous TPU v5p accelerators.

The key to this performance boost lies in the expansion of the TPU’s matrix multiple units (MXUs) and an increase in clock speed. Additionally, the Trillium TPU is equipped with 32GB of HBM operating at 1.6TB/s and a chip-to-chip interconnect that can reach 3.2 Tbps. This combination of enhancements is expected to significantly enhance the processing power and efficiency of AI models utilizing this new TPU.

RELATED POSTS

View all

view all