Google has taken a revolutionary step in artificial intelligence infrastructure and announced the new generation of special TPU (Tensoor Processing Unit) architecture. Ironwood, the seventh generation TPU of the company, has been designed to respond to the advanced needs of Gemini models. Google says that this processor makes advanced level tasks such as simulated reasoning. The technology giant calls this period as ım the inference age ”.
Google is talking about the new generation Gemini models, not only the model architecture, but also the infrastructure that makes the performance of these models possible. In the center of this infrastructure, Ironwood is now located. Together with this processor, the company states that “agent -like” artificial intelligence systems will gain strength. Such systems are defined as structures that can collect data on behalf of the user, make decisions and produce outputs. In short, Google announces that artificial intelligence evolves not only responding, but to a proactive structure.
Ironwood has been introduced as the most powerful and scalable TPU. Each chip has a transaction capacity up to 4,614 TFLOPS per second. These chips can communicate directly through the new generation Inter-Chip Interconnect (ICI) developed by Google. Moreover, these processors can work by clustering up to 9.216 pieces with their liquid cooling structures. This giant structure can reach a total of 42.5 Exaflops calculation power. Developers will be able to use these new generation processors in two different configurations: 256 chip servers or 9.216 chip clusters running with full capacity.
Ironwood chips attract attention not only with processing power, but also with memory capacity. Each TPU is equipped with 192 GB memory. This means a 6 -storey increase compared to the previous generation Trillium (TPU V6). The memory bandwidth is also improved by 7.2 TBPS 4.5 times. In addition, the company emphasizes that Ironwood is twice as efficient than Trillium in terms of performance per Watt.
Even Google is so confident in Ironwood that it doesn’t even refrain from making some controversial comparisons. The company claims that Ironwood is 24 times faster than El Capitan, known as the world’s most powerful super computer. Google does this comparison over the FP8, but El Capitan offers 1.7 exaflops.
Google’s Ironwood shows that a recent trend will continue unabated: independent of Nvidia. Although Nvidia is still undisputed in the sector, Amazon’s Trainium, Inferentia and Graviton chips are; Microsoft’s solutions like Maia 100, and finally Ironwood shows that technology giants attach great importance to internal solutions.
