At the Google I/O 2024 developer conference on Tuesday, Google unveiled its 6th generation custom chip, the Trillium TPU, which is scheduled to hit the market later this year, according to the report by TechCrunch.
According to the information provided by Google on its website, compared to TPU v5e, Trillium boasts a 4.7x peak compute performance increase per chip. Google has also doubled the High Bandwidth Memory (HBM) capacity and bandwidth, along with a 1x increase in Interchip Interconnect (ICI) bandwidth between chips.
Additionally, Trillium features the third-generation SparseCore, a dedicated accelerator for processing large embeddings, aimed at handling advanced ranking and recommendation workloads. Moreover, Trillium achieves a 67% higher energy efficiency compared to TPU v5e.
Trillium has the capacity to expand up to 256 TPUs within a singular pod boasting high bandwidth and low latency. Additionally, it incorporates multislice technology, allowing Google to interlink thousands of chips, thus constructing a supercomputer capable of facilitating a data center network capable of processing petabits of data per second.
In addition to Google, major cloud players such as AWS, Meta, and Microsoft have also made their way to develop their own AI Chips.
In late 2023, Microsoft unveiled two custom-designed chips, the Microsoft Azure Maia AI Accelerator, optimized for AI tasks and generative AI, and the Microsoft Azure Cobalt CPU, an Arm-based processor tailored to run general purpose compute workloads on the Microsoft Cloud. The former is reportedly to be manufactured using TSMC’s 5nm process.
In May 2023, Meta also unveiled the Meta Training and Inference Accelerator (MTIA) v1, its first-generation AI inference accelerator designed in-house with Meta’s AI workloads in mind.
AWS has also jumped into the AI chip market. In November, 2023, AWS released Trainium2, a chip for training AI models.
Read more
(Photo credit: Google)