Trillium, the Most Powerful and Energy-Efficient AI Chip Developed by Google to Date

Trillium, the Most Powerful and Energy-Efficient AI Chip Developed by Google to Date
Trillium, the Most Powerful and Energy-Efficient AI Chip Developed by Google to Date

Alphabet, Google’s parent company, has unveiled its latest addition to the AI ​​data center chip family: Trillium.

Innovation in AI Hardware

Representing the most advanced AI-specific hardware, Trillium is composed of Tensor Processing Units (TPUs). These custom AI data center chips are a standout alternative to Nvidia’s offerings, providing a compelling option in the market.

Currently, Nvidia owns 80% of the AI ​​chip market, with Google dominating the remaining 20%. It is notable that Google does not sell the chips, but instead rents them through its cloud computing platform.

Performance and Energy Efficiency

Trillium is the sixth generation of TPU, the most performant and energy efficient to date. Google has highlighted that Trillium TPUs achieve an impressive 4.7x increase in maximum compute performance per chip compared to the TPU v5e, as noted in a blog post.

The company’s latest offering features double the High Capacity Memory (HBM) capacity and bandwidth, along with double the Inter-Chip Interconnect (ICI) bandwidth. English) compared to TPU v5e.

Additionally, Trillium introduces a third generation of SparseCore, a specialized accelerator for processing ultra-large embeddings found in advanced classification and recommendation workloads. The blog post also highlights the ability to train the next wave of fundamental models at lower latency and lower cost.

Scalability and Efficiency

The latest Trillium TPUs stand out for being more than 67% more energy efficient than the v5e TPU, according to Google. Additionally, the Trillium is capable of scaling up to 256 TPUs within a single high-capacity, low-latency pod.

The blog also mentioned that beyond this pod-level scalability, Trillium TPUs, equipped with multislice technology and Titanium Intelligence Processing Units (IPUs), can scale to hundreds of pods, connecting dozens of thousands of chips in a building-scale supercomputer interconnected by a multi-petabits per second data center network.

Performance Improvements

The company achieved a 4.7x increase in compute performance per Trillium chip by increasing the size of the matrix multiplication units (MXUs) and increasing the clock speed.

In a blog post, the company stated: “Trillium TPUs will power the next wave of AI models and agents, and we look forward to helping enable our customers with these advanced capabilities.”

Impact on Cloud Computing Services

This advancement will greatly benefit Google’s cloud computing services and Gemini. Companies like Deep Genomics and Deloitte, which rely on Google Cloud services, will see a significant boost thanks to the new chip.

Support for training and serving long-context multimodal models on Trillium TPUs will enable Google DeepMind to train and serve future generations of Gemini models faster, more efficiently, and with lower latency than ever before.

Trillium TPUs are central to the Google Cloud AI Supercomputer, a supercomputing architecture designed specifically for edge AI workloads.

Gemini 1.5 Pro is Google’s largest and most capable AI model, and was trained using tens of thousands of TPU accelerators.

Our team is excited about the announcement of the sixth generation of TPUs, and we look forward to the increase in performance and efficiency for training and inference at the scale of our Gemini models.

Jeff Dean, chief scientist at Google Deepmind and Google Research.

Via Google com

If you liked this article, share it with your friends on social networks. Thank you!

 
For Latest Updates Follow us on Google News
 

-

PREV Kimbal Musk, Elon’s eccentric brother, who wears a cowboy hat and occupies a power chair at Tesla
NEXT I call for BQUO, the project to accelerate talent in DOMINION