According to reports, Google is planning to collaborate with MediaTek to develop its next-generation Tensor Processing Unit (TPU), an AI chip used internally by the company.
Based on information from *The Information* citing internal sources, the next-generation TPU might be manufactured by MediaTek at some point next year. The report also highlights that the production approach may differ, with MediaTek handling the input/output modules responsible for managing communication between the main processor and peripheral components.
Currently, Google’s TPUs are produced by Broadcom. Although MediaTek may take over the production of the upcoming chips, it is noted that Google has not yet ended its partnership with Broadcom.
Google’s TPU is a specialized application-specific integrated circuit designed to accelerate machine learning tasks, particularly those involving neural networks. These chips are optimized for Google’s TensorFlow framework, enhancing both training and inference processes by efficiently managing the computational demands of deep learning models.
Unlike traditional processors, TPUs are tailored for high-volume, low-precision arithmetic operations. This specialization gives TPUs a significant edge in performance and energy efficiency compared to general-purpose CPUs and GPUs.
In cloud environments, TPUs are integrated into Google’s data centers, providing scalable and efficient resources for large-scale machine learning workloads. For edge computing, Google offers the Edge TPU, a compact and energy-efficient version aimed at bringing AI capabilities to devices such as smartphones and IoT applications.
While TPUs have a variety of applications, their most notable use lies in AI. These dedicated processors are designed to accelerate machine learning tasks by efficiently handling the extensive computations required for training and inference, especially in neural network-based operations.
As Reuters pointed out, the ability to produce proprietary AI chips provides Google with a competitive advantage in the AI race, reducing reliance on Nvidia, which currently dominates the AI chip market.
The latest version of Google’s TPU is the sixth generation, known as the Trillium TPU, announced in October. As an alternative to Nvidia's popular GPUs, Trillium delivers four times the performance boost for AI training and three times the throughput for inference compared to its predecessor.
The Trillium TPU also features increased memory and bandwidth, enabling the chip to run larger language models with more weights and expanded key-value caches. It supports a broader range of model architectures, whether during training or inference.
