Ironwood: Google's New AI Accelerator Chip

Google Unveils Next-Generation TPU: Ironwood

At this week’s Cloud Next conference, Google introduced its newest AI accelerator chip, representing a significant advancement in its Tensor Processing Unit (TPU) line.

Introducing Ironwood

The newly developed chip, named Ironwood, marks Google’s seventh generation of TPUs. Notably, it is the first TPU specifically engineered for inference – the process of utilizing trained AI models.

Google Cloud customers can anticipate the availability of Ironwood later this year. It will be offered in two distinct configurations: a 256-chip cluster and a larger, more powerful 9,216-chip cluster.

According to Google Cloud VP Amin Vahdat, “Ironwood is our most powerful, capable, and energy-efficient TPU to date.” He further emphasized its design as being specifically tailored for large-scale inferential AI models.

Competitive Landscape

The arrival of Ironwood coincides with increasing competition within the AI accelerator market. While Nvidia currently holds a leading position, other major technology companies are actively developing their own in-house solutions.

Amazon offers its Trainium, Inferentia, and Graviton processors through AWS. Similarly, Microsoft provides Azure instances featuring its Maia 100 AI chip.

Performance and Architecture

Internal benchmarks conducted by Google reveal that Ironwood is capable of delivering peak computing power of 4,614 TFLOPs. Each individual chip incorporates 192GB of dedicated RAM, boasting a bandwidth approaching 7.4 Tbps.

A key feature of Ironwood is its enhanced SparseCore, a specialized core designed to efficiently process data commonly found in workloads such as “advanced ranking” and “recommendation” systems. For example, this could power algorithms suggesting products a user might be interested in.

Google’s architectural design prioritizes minimizing data movement and reducing latency within the chip itself. This optimization contributes to significant power savings.

Future Integration

Google intends to integrate Ironwood with its AI Hypercomputer, a modular computing cluster available within Google Cloud, in the coming months, as stated by Vahdat.

Vahdat concluded, “Ironwood represents a unique breakthrough in the age of inference,” highlighting its advancements in computation power, memory capacity, networking capabilities, and overall reliability.

Note: This article was updated at 10:45 a.m. Pacific time to correct a previous misidentification of Microsoft’s Cobalt 100. Cobalt 100 is a general-purpose chip, while the Maia 100 is specifically designed as an AI chip.