Driven by its latest self-developed TPU (Tensor Processing Unit) chip, tech giant Google is moving towards AI ( Artificial Intelligence) . Nvidia, the chip leader Launch an attack.
Recently, research firm SemiAnalysis released a report explaining how Google has become Nvidia's most threatening competitor in the commercial chip market. SemiAnalysis' model data shows that, from a TCO (Total Cost of Ownership) perspective, Google's seventh-generation TPU chip has an overwhelming advantage over Nvidia in cost efficiency: the more Google TPUs customers purchase, the more Nvidia can save in GPU capital expenditure.
SemiAnalysis points out that, taking OpenAI as an example, since the release of its large model GPT-4o in May 2024, OpenAI has not yet completed any large-scale pre-trained deployments for "next-generation frontier models." Although OpenAI has not yet switched from Nvidia GPUs to Google TPUs, under the competitive pressure brought by TPUs, OpenAI has already obtained a discount of about 30% on its computing cluster quotes from Nvidia.
TPU (Tensor Processing Unit) is a self-developed AI-specific integrated circuit (ASIC) unveiled by Google in 2016, primarily used to accelerate the training and inference of machine learning models. Compared to Nvidia's GPUs, TPUs have less versatility, are more specialized, but consume less power. Previously, Google did not sell TPUs directly, but instead allowed other companies to rent them through Google Cloud.
SemiAnalysis emphasizes that Google's AI infrastructure advantage does not lie in chip specifications, but rather in its system-level engineering capabilities, meaning that "the system is more important than the microarchitecture": "By relying on system design, interconnects, compilers, and overall hardware and software synergy, Google is able to make its TPUs comparable to Nvidia's in terms of actual performance and cost efficiency."
In November of this year, Google announced the imminent release of its seventh-generation self-developed TPU chip, Ironwood. According to reports, Ironwood offers more than four times the performance in training and inference compared to its predecessor. It employs an advanced interconnect architecture, allowing a single POD (AI supercomputer unit) to connect up to 9,216 Ironwood chips, forming a "super pod" with a breakthrough ICI (inter-chip interconnect network) capable of 9.6 Tb (terabits) per second, accessing up to 1.77 PB of shared high-bandwidth memory. (HBM).
According to data from SemiAnalysis, from a technical perspective, the Ironwood chip nearly matches Nvidia's Blackwell GPU series in terms of floating-point operations per second (FLOPs) and memory bandwidth. If Google were to use this chip, the total cost of ownership (TCO) per chip would be approximately 44% lower than that of a comparable Nvidia GB200 system.
SemiAnalysis believes that Google is striving to become a "truly differentiated cloud service provider." In October, AI startup Anthropic announced a partnership with Google to deploy up to one million Google TPU chips to train its large AI models. Of the agreed-upon one million chips, the first batch of approximately 400,000 of the latest TPU chips will be manufactured by Broadcom , which co-designed the chips with Google. The chips were sold directly to Anthropic for approximately $10 billion; the remaining 600,000 chips were leased through Google Cloud.
Currently, Nvidia is also feeling the pressure from Google. On November 25th local time, reports surfaced that Meta is considering implementing measures in its data centers starting in 2027. The deployment of Google's TPUs (Tensor Processing Units), worth billions of dollars, and the potential leasing of TPUs through Google Cloud next year, caused Nvidia's stock price to fall by more than 7% during the day.
In response, Nvidia stated publicly that its GPUs "remain a generation ahead of the industry": "Nvidia offers higher performance, greater versatility, and better substitutability compared to ASICs (Application-Specific Integrated Circuits) designed specifically for a particular AI framework or function."
(Source: The Paper)
