Google Launches 7th-Generation Ironwood AI Accelerator Chip Cluster, 24 Times Faster Than Top Supercomputer

Google has introduced its revolutionary “7th-generation”custom AI accelerator named Ironwood, marking a significant milestone as the company’s first in-house chip explicitly designed for inference workloads.

The Breakthrough of Google’s Ironwood AI Chip: Setting New Standards for Inference Workloads

Unveiled during the Google Cloud Next 25, the Ironwood accelerator stands out as Google’s most powerful and efficient chip to date. With enhanced generational capabilities, it showcases the company’s commitment to advancing inference workloads, which Google believes represents the next crucial phase of artificial intelligence. This cutting-edge accelerator will be available in two configurations for Google Cloud customers: a 256-chip setup and a more robust 9, 216-chip setup designed to cater to various workload and inference power requirements.

The Ironwood AI chip marks a transformative step for contemporary AI markets. Under the 9, 216-chip configuration, the system reportedly achieves 24 times the computing power of the esteemed supercomputer El Capitan, reaching an impressive 42.5 Exaflops. Further emphasizing the advancements made, Ironwood also boasts a doubling of performance per watt when compared to the previous-generation Trillium TPU, highlighting the significant upward trajectory of performance enhancements over generations.

Substantial increase in High Bandwidth Memory (HBM) capacity: Each Ironwood chip features 192 GB, a sixfold increase compared to Trillium. This expansion allows for the processing of larger models and datasets, minimizing the frequency of data transfers and enhancing overall performance.

Dramatically improved HBM bandwidth: Ironwood reaches bandwidth levels of 7.2 TBps per chip, marking a 4.5-fold improvement over Trillium’s bandwidth. This rapid data access is essential for the memory-intensive tasks prevalent in modern AI applications.

Enhanced Inter-Chip Interconnect (ICI) bandwidth: The ICI bandwidth has seen a significant increase to 1.2 Tbps bidirectional, an improvement of 1.5 times compared to Trillium. This enhancement facilitates quicker communication between chips, enabling efficient distributed training and inference at scale.

The launch of Ironwood underscores the impressive progress achieved in custom in-house AI solutions, positioning Google as a formidable contender in a market traditionally dominated by NVIDIA. Performance metrics indicate a vast potential for growth, especially as competitors like Microsoft and their Maia 100, and Amazon with their Graviton chips, recognize and seize the opportunities available through tailored AI solutions.

Introducing 7th Generation TPUs: Ironwood (https://www.youtube.com/watch?v=9i1ZM0dPyRo)

Source&Images