Google presents Ironwood, new chip for AI that can rival gpus nvidia

Google officially launched this week Ironwood, the seventh generation of its AI accelerator chips. The big differential of the new release is that this is the first product of Google with a focus on inference, according to the company.

Inference is the specific name of the most important process of the Generative AI, which has currently been popularized with bots such as Google’s own chatgpt or Gemini. The company says that “Ironwood is our most powerful, capable and efficient TPU so far, created for strong reasoning in scale inferential AI models.”

TPU is an acronym for Tensor Processing Unit, as Google calls its accelerators to AI. According to Reuters, Ironwood possibly arrives as one of the few viable options to face Nvidia GPUS, which have dominated the professional AI segment.

Ironwood, however, has its use somewhat limited. The chip is focused on internal use by Google’s own engineers, or to customers who hire their cloud services. The company’s Gemini is processed entirely by its chips, so you should receive a new boost with the arrival of the new generation.

Ironwood will be implemented in settings of up to 9,216 chips

Google states that the new generation of its accelerator can be climbing up to 9,216 chips with liquid cooling, connected by an advanced ICI (Inter chip interconnect) network.

Thus, the company will offer Google Cloud services in two settings: 256 POD chips and 9,216 chips by POD.

According to the company, the most advanced option equipped with the Ironwood components delivers a total of 42.5 Examflops, which would be up to 24 times more computational power than El Capitan offers. The current largest supercomputer in the world delivers 1.7 Examflop by Pod.

Related News:

Google has not reported, however, which semiconductor manufacturer is producing the new Ironwood.