Google I/O 2024 once again showed that Google is focussing heavily on artificial intelligence. In addition to numerous software announcements, the focus was also on hardware innovations and infrastructure improvements that have the potential to significantly change the AI landscape.
Trillium: the sixth generation of TPU
One of the most significant announcements was the presentation of Trillium, the sixth generation of Google’s Tensor Processing Units (TPUs). These chips, specially developed for AI workloads, are designed to deliver an impressive increase in performance. With up to 4.7 times higher peak performance and double the memory bandwidth compared to the previous generation TPU v5e, Trillium promises to train and execute sophisticated AI models even more efficiently. Trillium is also more energy efficient by over 67%, which makes an important contribution to sustainability.
Axion: Google’s first ARM-based chip
Another milestone is the Axion processor, Google’s first self-developed chip based on the Neoverse V2 design of the ARM architecture. This chip was specially developed for the Google Cloud data centres and is intended to provide a significant increase in performance. According to Google, Axion offers up to 30% more performance than the fastest general-purpose ARM-based instances in the cloud and is up to 50% faster and 60% more energy-efficient than comparable x86-based instances.
NVIDIA Blackwell: Expansion of the partnership for more AI power
Google Cloud is growing its partnership with NVIDIA to integrate the new Blackwell GPU architecture into its cloud platform. Blackwell promises a significant increase in performance for AI and high-performance computing applications and will be used in various areas in the coming years. Integration into Google Cloud will enable customers to utilise Blackwell’s performance for their own AI projects. Initial benchmarks indicate an up to 30-fold increase in inferencing performance and up to 25-fold improvement in energy efficiency for large language models (LLMs) compared to conventional GPUs. In addition, the 5th generation of NVLink enables GPU-to-GPU communication at up to 1.8 TB/sec, which is twice as fast as the 4th generation used by Hopper.
AI hypercomputers and sustainable infrastructure
Google also emphasised its progress in the area of AI infrastructure at its I/O conference. The AI Hypercomputer, a supercomputing architecture for AI, combines CPUs, GPUs and TPUs to achieve twice the efficiency of conventional hardware and chips. This impressive increase in performance is made possible by a combination of different factors:
- Hardware: high-performance CPUs, GPUs and TPUs at the heart of the AI hypercomputer.
- Open source software: The use of open source frameworks such as SAX, TensorFlow, PyTorch and Kubernetes Engine accelerates and simplifies the development and deployment of AI applications.
- Flexible utilisation: Dynamic workload schedulers, CUDA and on-demand spot instances enable efficient use of resources and optimal adaptation to the respective requirements.
- Liquid Cooling: With a capacity of one gigawatt, the innovative cooling technology contributes to high energy efficiency by effectively cooling the high-performance hardware. However, it should be noted that this technology involves high water consumption, an aspect that should not be ignored in the discussion about sustainability (keyword: Green IT).
The infrastructure also includes Google Cloud’s fibre optic network, which is almost ten times longer than that of the next largest cloud provider at 2 million miles. This enables fast and reliable connections for AI applications and other cloud services.
Outlook
The hardware innovations and infrastructure improvements presented at Google I/O 2024 show that Google continues to invest heavily in the development of AI-specific hardware and infrastructure. With Trillium, Axion, the expanded partnership with NVIDIA for Blackwell and the powerful AI hypercomputer, Google is positioning itself as a leading provider of AI infrastructure and enabling companies and developers to take full advantage of the benefits of artificial intelligence. It will be interesting to see what further developments in this area will follow in the near future.