Google announced during its annual developer conference the next generation of Tensor Processing Units, which are artificial intelligence chips for data centers.
The release of this sixth generation of chips, known as Trillium, is expected later this year.
These Tensor Processing Units are similar to Google’s GPUs and Nvidia’s graphics processing units, with the latter being commonly used for running applied artificial intelligence workloads.
Google aims with Trillium to provide an alternative to Nvidia’s GPUs for businesses.
Sundar Pichai, CEO of Google, stated: “We have been at the forefront of GPUs for over a decade.”
The announcement of the next generation Tensor Processing Units is a significant step at the company’s annual developer conference, even though the chips will be released later in the year.
Upon arrival, the chip stands out with up to 4.7 times performance increase in maximum computing performance per chip compared to the previous generation.
Google has partially achieved this by expanding “chip matrix multiplication” units and increasing speed. Additionally, Google has also increased the memory bandwidth for Trillium chips.
Trillium chips feature third-generation SparseCore technology, specifically developed to accelerate processing of massive embedding operations that occur in advanced workloads like classification and recommendations.
The company emphasizes that this helps Trillium units to train models faster and deliver their services with lower latency.
Pichai described the new chips as the most energy-efficient from Google to date, which is important given the increasing demand for artificial intelligence chips.
Pichai stated that the demand for machine learning technologies has increased by a million times over the past six years, equivalent to almost a tenfold increase annually.
This growth is not sustainable without investment in reducing the energy requirements for these chips. According to Google, the new Trillium units are 67% more energy-efficient compared to fifth-generation chips.
Google’s Tensor Processing units tend to appear in multiple releases. So far, the company has not provided any additional information about the new chips or their cost for Google Cloud services.
Earlier this year, the company announced that they will be at the forefront of cloud service providers offering access to next-generation Blackwell processors from Nvidia.
This means that developers will have to wait until early 2025 to get these chips.
Pichai said they will continue to invest in the infrastructure to support their advancements in artificial intelligence and explore new horizons.
