The latest TPU and the upcoming Ironwood supercomputer were just the start
Google is taking the next step in its quest to become a serious challenger to Nvidia GPUs. As I recently noted in my post on CSP silicon costs and failures, Google is the exception. Google possesses the silicon, software, and systems expertise to challenge Nvidia, at least in its own cloud.
Now, the Information has reported that Google has approached Neocloud companies, offering TPUs to be hosted on-site for Neocloud customers. This is a critical but inevitable shift in Google’s compute strategy. Google would rather compete with the smaller cloud service providers using their silicon than using Nvidia.
At the recent HotChips conference, Google presented more details about the massive AI supercomputer they are building, with availability expected later this year. Ironwood is AI turned up to 11, featuring over 9,000 next-gen TPUs interconnected by a flexible optical fabric that enables the dynamic allocation of TPU resources.
Google may be talking about dropping TPU7 (Ironwood) pods into these data centers. These would occupy space and power that is likely intended for Nvidia GPU deployments. The TPU7X supercomputer described above is for Google Cloud Compute deployment.
As Nvidia’s recent stock pullback indicates, the competition has arrived. I don’t think it’s China, at least not yet. It’s Google and Amazon AWS.