Google has announced the launch of two new custom-designed artificial intelligence chips, marking a significant escalation in its efforts to compete with industry leader Nvidia. The new processors, named the Axion CPU and the sixth-generation Tensor Processing Unit (TPU) called Trillium, were unveiled at the Google Cloud Next conference. This move signals Google’s intent to reduce reliance on external suppliers and offer more cost-effective and powerful AI infrastructure to its cloud customers.
Performance and Cost Improvements
According to Google, the latest TPUs deliver substantial performance gains over their predecessors. The company stated the Trillium TPUs are roughly 4.7 times faster in compute performance per chip compared to the previous TPU v5e generation. Furthermore, Google claims the new chips offer improved energy efficiency. The Axion CPU, based on Arm’s Neoverse 2 design, is presented as a high-performance central processor for general computing workloads within the Google Cloud ecosystem.
The primary value proposition for customers is a combination of increased speed and reduced cost. Google emphasized that these new in-house processors provide a more economical path for running large-scale AI training and inference workloads compared to utilizing hardware from other vendors. This development is part of a broader industry trend where major cloud providers are developing proprietary silicon to gain control over their supply chains and optimize performance for their specific software stacks.
A Continued Partnership with Nvidia
Despite launching competitive products, Google clarified that it remains a committed partner to Nvidia. The company confirmed it will continue to offer Nvidia’s latest GPUs, including the Blackwell architecture-based systems, on its Google Cloud platform. This dual-strategy approach allows Google to provide customers with a choice between its own optimized hardware and the industry-standard platforms from Nvidia, catering to different technical requirements and preferences.
Industry analysts note that this is a common tactic in the competitive cloud market. By developing custom chips, Google aims to capture more value from its cloud services and differentiate its offerings. However, maintaining support for Nvidia’s hardware is considered essential, as much of the global AI software ecosystem is built and optimized for Nvidia’s CUDA platform, creating a significant barrier to switching for many developers and enterprises.
Market Context and Implications
The announcement places Google in direct competition with other cloud giants pursuing similar strategies. Amazon Web Services has its Graviton CPUs and Trainium/Inferentia AI chips, while Microsoft Azure has introduced its Maia AI accelerator and Cobalt CPU. This collective move by hyperscalers challenges Nvidia’s near-monopoly in the data center AI accelerator market, potentially leading to more innovation and price competition.
For businesses and developers, the proliferation of custom silicon options could lead to more tailored and cost-efficient cloud computing solutions. However, it also raises considerations about software portability and vendor lock-in, as applications optimized for one provider’s custom hardware may not run as efficiently on another’s.
Looking ahead, Google plans to make the Trillium TPUs available to cloud customers later in 2024. The company’s continued investment in custom silicon research and development suggests further iterations and more specialized processors can be expected. The long-term impact on Nvidia’s market share will depend on the adoption rate of Google’s new chips and the ability of the software community to adapt to the expanding array of hardware architectures available in the cloud.
Source: GeekWire
