Ironwood, the key name of the New processor for Googlehas been the Hardware star of the Next 25 event where Google Cloud has presented its novelties in artificial intelligence solutions for companies.
Ironwood is the seventh generation of TPU (tensioning processing units) that Google designs internally for Accelerate automatic learning tasks of its most powerful Gemini models. Google’s most powerful to date, the company states that this chip represents an important change that will unk “Era of inference”.
Whenever Google talks about the capacities of a new version of Gemini, he points out that they are linked not only to the code, but also to their infrastructure. Your custom AI hardware is a key element to accelerate inference and expand context windows.
With Ironwood, Google says Your most scalable and powerful TPU to datewhich will allow AI to act on the name of the user to collect data proactively and generate results. This refers to Google when he talks about the agent.
The new chip offers greater performance compared to the previous tensioning processing units, since it is designed to operate in groups of up to 9216 chipswhich will communicate directly with each other through an interconnection between recently renewed chips (ICI) and will be ventilated by a liquid cooling system to keep temperatures at bay.
Google says that this type of design will be a great advantage not only to feed their great Gemini language models, but also for developers who seek to execute AI projects in the cloud. Developers may take advantage of Ironwood in two different configurations: a 256 chips server or the 9216 chips complete cluster.
In its largest version, Google’s Ironwood pods can generate the amazing amount of 42.5 Inference computing ex -Flops. Each chip has a maximum performance of 4614 TFLOP, which, according to Google, is a substantial improvement with respect to the previous chips. Google has also increased the memory of the new TPU, with 192 GB each, six times more than the latest generation Trillium TPU of Google. Memory bandwidth has also increased to 7.2 Tbps, with an improvement of 4.5 times.
It is not easy to place the precise performance of the new Google chip since there are numerous ways to measure the performance of AI accelerators, which hinders comparison between them. Google uses FP8 precision as a reference for the new TPU, but compares it with some systems, such as the El Capitan supercomputer, which do not admit FP8 in hardware. Therefore, it is convenient to take its statement that the Ironwood modules are “24 times faster” what comparable segments of the most powerful supercomputer in the world.
