At its next Cloud Next conference this week, Google unveiled the latest generation of its TPU AI accelerator chip.
The new chip, called Ironwood, is the TPU of the seventh generation of Google and is the first optimized for inference-that is to say, execute models of AI. Planned to launch a little later this year for Google Cloud customers, Ironwood will be available in two configurations: a 256 chip cluster and a 9,216 chip cluster.
“Ironwood is our most powerful, the most capable and the most energy-efficient TPU,” wrote Google Cloud Amin Vahdat vice-president in a blog article provided in Techcrunch. “And it is specially designed for the thought of power, large -scale inferential AI models.”
Ironwood arrives while competition in the ACA accelerator is warmed up. Nvidia can have their heads, but technological giants, including Amazon and Microsoft, push their own internal solutions. Amazon has its Trainium, Inferentia and Graviton processors, available via AWS, and Microsoft hosts Azure instances for its Maia 100 AI chip.
Ironwood can deliver 4,614 Tflop of calculation power to the peak, according to Google’s internal comparative analysis. Each chip A 192 GB of dedicated RAM with a bandwidth approaching 7.4 TBPS.
Ironwood has an improved specialized nucleus, sparsecore, for the processing of common data types in the workloads “advanced classification” and “recommendation” (for example, an algorithm that suggests clothes you might like). TPU architecture has been designed to minimize data movement and chip latency, leading to power savings, says Google.
Google plans to integrate Ironwood with its hypercombummer AI, a modular computer cluster in Google Cloud, in the near future, added Vahdat.
“Ironwood represents a unique breakthrough in the era of inference,” said Vahdat, “with increased computing power, memory capacity, … Networking progress and reliability.”
Update 10:45 Pacific: an earlier version of this lower story to cobalt 100 from Microsoft as a AI chip. In fact, Cobalt 100 is a chip for general use; Maia 100 from Microsoft is an AI chip. We have corrected the reference.
