Google’s seventh-generation Tensor Processing Unit (TPU) is set to be unleashed in the coming weeks, offering staggering throughput speeds when interconnected.

Dubbed Ironwood, Google’s latest custom hardware was unveiled back in April and is specifically designed with inference in mind, with each individual chip offering peak compute of around 4,600 teraflops.

Using its proprietary networking tech, Inter-Chip Interconnect (ICI), the tech giant plans to combine 9,216 chips into a single pod that’s capable of 42.5 exaflops – that’s 24x more computing power than the world’s most powerful supercomputer, El Capitan, which only offers 1.742 exaflops per pod.

It’s important to note, however, that El Capitan’s performance is measured in FP64 precision. Google’s Ironwood pods, meanwhile, use FP8 precision benchmarking, or 8-bit precision calculations, to calculate the compute performance of Ironwood. While AI performance is measured in FP8, traditional compute performance is measured in double-precision calculations, also known as FP64. As a result, a system boasting 100 petaflops of FP64 performance is more powerful than a system that has achieved 100 petaflops of FP8 performance, for example.

Google’s networking tech enables remote direct memory access (RDMA) over PCIe P2P, meaning each TPU can be interconnected and communicate with another with 400G bandwidth going in each direction.

The tech giant likens ICI and its capabilities to turn its latest TPUs into “one brain.”

Where traditional supercomputing stacks are limited by flow issues, causing what are essentially data traffic jams, the hyperscaler claims its proprietary ICI tech can support fast enough speeds to keep workloads moving.

“This unprecedented speed ensures that the data being worked on by one chip is instantly available to the 9,215 other chips, allowing them to work together with minimal delay,” a company blog post reads.

Further supporting the enhanced flow of data between TPUs is Optical Circuit Switching (OCS), another Google technology that routes around interruptions to keep workloads running.

Among the early adopters of Ironwood-based computing power is Anthropic, the AI startup taking on OpenAI with Claude.

James Bradbury, Anthropic’s head of compute, said Ironwood’s improvements across inference performance and training scalability “will help us scale efficiently while maintaining the speed and reliability our customers expect.”

Google’s Ironwood TPUs are yet another powerful computing system available to Anthropic. The Claude developer has free rein of Project Rainier, the mammoth cluster from Amazon Web Services, which went live last week. Like Amazon, Google has also invested in Anthropic.