Trillium, the sixth iteration of Google’s Tensor Processing Unit (TPU), is nearly five times more efficient than its predecessor, TPUv5, in peak compute performance and memory bandwidth, Google said.
Google unveiled a new chip, Trillium, for training and running foundation large language models such as Gemma and Gemini at its annual I/O conference on Tuesday.
Trillium is the sixth iteration of Google’s Tensor Processing Unit (TPU) and is 67% more energy efficient and nearly five times as fast as its predecessor, TPU v5, according to the company. Google plans to use Trillium in its AI Hypercomputer, a supercomputing architecture designed for cutting edge AI-related workloads, and will make the chips available to enterprises by the end of the year.
“Trillium TPUs achieve an impressive 4.7X increase in peak compute performance per chip compared to TPU v5e. We doubled the High Bandwidth Memory (HBM) capacity and bandwidth, and also doubled the Interchip Interconnect (ICI) bandwidth over TPU v5e,” Amin Vahdat, general manager of systems and cloud AI at Google, wrote in a blog post.
The increase in compute performance, according to Vahdat, is achieved by expanding the size of matrix multiply units (MXUs) and increasing the clock speed, which in turn makes it possible to train the next wave of foundation models faster and run them with reduced latency and lower cost.
MXUs are part of the TPU chip architecture. Typically, a TPU chip contains one or more TensorCores and each of these TensorCore consists of one or more MXUs, a vector unit, and a scalar unit.
Trillium chips can scale up to 256 TPUs in a single high-bandwidth, low-latency pod, Vahdat added.
Other Trillium features include dataflow processors that accelerate models relying on embeddings found in recommendation models, and support for more high-bandwidth memory (HBM) in order to work with larger models with more weights and larger key-value caches.
More slices
Further, Trillium comes with Google’s multislice technology, which the company introduced for the first time, in preview, while unveiling TPU v5e last year in August.
Multislice technology, according to the company, allows enterprise users to easily scale AI models beyond the boundaries of physical TPU pods — up to tens of thousands of Cloud TPU v5e or TPU v4 chips.
Before the release of this technology, training jobs using TPUs were limited to a single slice of TPU chips, capping the size of the largest jobs at a maximum slice size of 3,072 chips for TPU v4.
“With Multislice, developers can scale workloads up to tens of thousands of chips over inter-chip interconnect (ICI) within a single pod, or across multiple pods over a data center network,” Vahdat explained last year in a blog post co-written with his colleague Mark Lohmeyer.
Open source support
Trillium will support open source libraries, such as JAX, PyTorch/ XLA, and Keras 3, Vahdat said. “Support for JAX and XLA means that declarative model description written for any previous generation of TPUs maps directly to the new hardware and network capabilities of Trillium TPUs,” he wrote, adding that Google has partnered with Hugging Face on Optimum-TPU for streamlined model training and serving.
Google launched the first iteration of its TPU in 2016.
Most hyperscalers, including the likes of Microsoft, AWS, and IBM, have started developing their own chips for AI workloads as they face demand on one hand and shortage of Nvidia GPUs on the other.
While AWS has been iterating on its Tranium and Inferentia accelerators, Microsoft, last year, released its Cobalt CPU and Maia accelerator chips.
