google launches the next generation of its custom ai chips

Google Unveils Next-Generation TPU v4 AI Chips

During its annual I/O developer conference, Google has revealed the latest iteration of its custom Tensor Processing Units (TPUs). These new AI chips represent the fourth generation in the TPU lineage.

According to Google, the performance of these v4 TPUs is doubled compared to their predecessors.

Enhanced Computing Power Through Pods

Google CEO Sundar Pichai highlighted that these chips are interconnected within pods, each containing 4,096 v4 TPUs.

A single pod configuration is capable of delivering computational performance exceeding one exaflop.

Availability on Google Cloud

While Google leverages these specialized chips to enhance its internal machine learning capabilities, they will also be offered to developers.

Access will be provided through the Google Cloud platform.

A Milestone in Computing

“This represents the fastest system we have ever implemented at Google, and marks a significant achievement for our company,” stated Pichai.

Previously, achieving an exaflop of computing power necessitated the construction of a dedicated supercomputer.

Google now has numerous TPUv4 pods already operational, with plans to deploy dozens more in its data centers.

Many of these pods will operate utilizing nearly 90% carbon-free energy sources.

The availability of TPUv4 pods for Google Cloud customers is scheduled for later this year.

Early Investment in Custom Hardware

The TPUs were among Google’s initial forays into custom chip development.

Unlike competitors like Microsoft, who opted for the adaptability of FPGAs for their machine learning services, Google strategically invested in these custom chips.

Trade-offs in Development and Performance

Developing custom chips requires a longer timeframe and faces the challenge of rapid technological obsolescence.

However, this approach can yield substantial gains in performance.