Google Unveils Trillium: Energy-Efficient Cloud TPU Boosts AI

May 16, 2024: Google Cloud has unveiled Trillium, its sixth-generation Tensor Processing Unit (TPU), marking a significant advancement in artificial intelligence hardware. Notably, Trillium boasts the distinction of being Google’s most energy-efficient TPU. This innovation is poised to empower the development and deployment of the next generation of AI models.

Trillium surpasses its predecessors in several key metrics. It delivers a remarkable 4.7-fold increase in peak compute performance per chip compared to the TPU v5e. This enhanced processing prowess is attributed to advancements in the chip’s design, including expanding its matrix multiply units and increasing its overall clock speed. Additionally, Trillium features double the memory bandwidth of its predecessor, enabling it to handle demanding workloads more efficiently.

Beyond raw performance, Trillium incorporates the third generation of SparseCore technology. This specialized accelerator is designed to expedite the processing of intricate data structures known as embeddings, which are prevalent in sophisticated ranking and recommendation algorithms. Incorporating SparseCore is expected to accelerate the training of next-generation AI models while concurrently reducing latency and lowering operational costs.

Furthermore, Google emphasizes Trillium’s exceptional energy efficiency. The new TPU is touted to be 67% more energy-efficient than the TPU v5e. This advancement is crucial in the ever-growing demand for AI processing power. The exponential growth of machine learning workloads necessitates the development of more sustainable hardware solutions, and Trillium represents a significant stride in this direction.

Trillium boasts scalability, offering the capability to be configured with up to 256 TPUs within a single, high-bandwidth, and low-latency pod. These pods can be further scaled into hundreds using Google’s multislice technology, unveiled in late 2023. This enables the interconnection of tens of thousands of chips via Google’s Jupiter data center network, facilitating the execution of massive-scale AI workloads.

The introduction of Trillium signifies Google’s continued commitment to pioneering advancements in AI hardware. This novel TPU is expected to empower researchers and developers to push the boundaries of what’s achievable in the field of artificial intelligence. It is anticipated that Trillium will play a pivotal role in the development and deployment of the next generation of AI models, fostering breakthroughs across various domains.

Also Read, Baird Shifts Squarespace (SQSP) Rating to Neutral