Google unveils Trillium chip, marking a quantum leap in AI data centre performance

Alphabet Inc, Google's parent company, has announced an advancement in artificial intelligence (AI) technology with the launch of its latest AI data centre chip, Trillium.

This sixth-generation chip promises to be nearly five times faster than its predecessor, bringing a boost in computational power and efficiency for AI applications.

In a press briefing on Tuesday, Alphabet CEO Sundar Pichai noted the explosive growth in demand for machine learning capabilities.

"Industry demand for machine learning computation has grown by a factor of 1 million in the last six years, roughly increasing 10-fold every year," Pichai said.

He highlighted Google's decade-long goal to perfect AI chip technology, positioning the company to meet this increasing demand.

The Trillium chip represents one of the few credible alternatives to Nvidia’s dominant processors in the AI data centre market.

Currently, Nvidia holds approximately 80 per cent of this market, with the majority of the remainder occupied by various iterations of Google's tensor processing units (TPUs).

Unlike Nvidia, Google does not sell its chips directly but offers access through its cloud computing platform, making these powerful processors available to a broad range of users.

According to Google, the Trillium chip delivers 4.7 times better computing performance compared to the TPU v5e, a chip designed for generating text and media from large AI models.

Additionally, the Trillium processor boasts a 67 per cent increase in energy efficiency over its predecessor, a crucial improvement in reducing operational costs and environmental impact.

The new chip, which will be available to Google's cloud customers in late 2024, achieves its performance gains through enhanced high-bandwidth memory capacity and overall bandwidth.

This advancement addresses a significant bottleneck in AI model performance, which relies heavily on vast amounts of advanced memory.

Engineers at Google have designed the Trillium chips to be deployed in pods of 256 chips, with scalability reaching up to hundreds of pods.

This modular approach allows for flexible and powerful configurations tailored to the needs of various AI workloads, enhancing the versatility and reach of Google's AI infrastructure.

(With inputs from Reuters)