Google Cloud has just announced the latest generation of artificial intelligence (AI) chips, marking the next step in its hardware self-development strategy to optimize performance and cost on the cloud platform.
Notably, this eighth-generation chip line is separated into two separate variants, serving two core needs of AI: training and reasoning.
Google introduces TPU 8t for model training, which is a stage requiring extremely high computing power to teach AI systems to understand data.
Meanwhile, TPU 8i is optimized for inference, that is, the modeling process learned is used to answer questions and process user requests in real time.
According to Google, the new TPUs bring a significant leap in performance. Compared to the previous generation, the model training speed can be up to 3 times faster.
Notably, the system also has the ability to connect more than 1 million TPUs in the same cluster, opening up extremely large computing scale for complex AI applications.
Besides performance, cost and energy factors are also emphasized. TPU chips designed by Google themselves are famous for their power saving ability, helping businesses significantly reduce operating costs when deploying AI on a large scale.
This is an important advantage in the context of increasing demand for AI, leading to pressure on infrastructure and electricity.
However, the launch of the new generation TPU does not mean that Google is turning its back on Nvidia (the giant dominating the global AI chip market).
In fact, Google affirms that TPUs will play an additional role, not replacing Nvidia's GPU systems in its cloud infrastructure.
The company even said it will integrate Nvidia's latest chips, including Vera Rubin architecture, into the service in the near future.
This trend is not only for Google. Large cloud providers such as Microsoft and Amazon are also developing internal AI chips to reduce third-party dependence and optimize costs.
However, in the short term, Nvidia still maintains an almost irreplaceable position.
Analyst Patrick Moorhead predicted in 2016 that Google's TPUs could be detrimental to Nvidia. But the reality now shows the opposite when Nvidia has risen to become a company with a market capitalization of nearly 5,000 billion USD, thanks to the AI boom.
Even the development of cloud platforms like Google Cloud may continue to benefit Nvidia. As AI demand increases, businesses will need more computing resources, including Nvidia's GPUs and Google's TPUs.
Not stopping there, the two giants also expanded their cooperation. Google said it is cooperating with Nvidia to optimize network performance for GPU systems on cloud platforms, through Falcon technology, which is a software network solution developed by Google and publicly source coded under the patronage of the Open Compute Project.
The combination of self-developing chips and strategic cooperation shows that Google is pursuing a balanced direction, both building independent capabilities and taking advantage of the existing ecosystem.
