Google Cloud Next is currently taking place in Las Vegas, showcasing a range of new instance types and accelerators for the Google Cloud Platform. Aside from the new custom Arm-based Axion chips, the majority of this year’s announcements revolve around AI accelerators, which are either developed by Google or sourced from Nvidia.
Recently, Nvidia made an announcement about its Blackwell platform. However, it is unlikely that Google will make those machines available in the near future. The inclusion of the high-performance Nvidia HGX B200 for AI and HPC workloads, as well as the GB200 NBL72 for large language model (LLM) training, is expected to be available in early 2025. One intriguing detail from Google’s announcement is that the GB200 servers will utilize liquid-cooling technology.
It should be noted that Nvidia has stated that its Blackwell chips will not be accessible to the general public until the final quarter of this year.
Prior to the establishment of Blackwell
Google has also introduced the A3 Mega instance, which provides developers with increased power for training LLMs. Developed in collaboration with Nvidia, this particular instance incorporates the widely accepted H100 GPUs. However, it also integrates a novel networking system capable of delivering double the bandwidth per GPU.
Google has introduced a new A3 instance called A3 Confidentiality. Google designed this instance to enhance the security of sensitive data and AI workloads during training and inferencing. The company has a history of providing secure computing services that encrypt data during use. In this case, once activated, confidential computing will ensure that data transfers between Intel’s CPU and the Nvidia H100 GPU are encrypted using protected PCIe. According to Google, no modifications to the code are necessary.
Google recently introduced its Cloud TPU v5p processors, which are the latest and most advanced AI accelerators developed in-house. The public can now widely access these processors. The chips demonstrate a significant increase in floating-point operations per second and memory bandwidth speed.
For optimal performance, high-speed processors require a compatible and efficient underlying architecture. In addition to the new chips, Google also announced on Tuesday new AI-optimized storage options. Google has introduced Hyperdisk ML, a new block storage service that is currently in preview. According to Google, this service can significantly enhance model load times, improving them by up to 3.7 times.
Google Cloud is introducing a range of conventional instances that utilize Intel’s fourth- and fifth-generation Xeon processors. The new general-purpose C4 and N4 instances will include the latest fifth-generation Emerald Rapids Xeons. While the N4 instance prioritizes affordability, the C4 instance prioritizes performance. The new C4 instances are currently in private preview, while the N4 machines are now generally available.
Additionally, there are new C3 bare-metal machines, which are currently in preview. These machines are equipped with older, fourth-generation Intel Xeons. The X4 memory-optimized bare metal instances are another preview offering. Lastly, Google Cloud has introduced the Z3, their first storage-optimized virtual machine. The Z3 is designed to provide the highest IOPs for storage-optimized instances compared to other leading cloud providers.
