Google Cloud and NVIDIA Collaborate on A3 Supercomputer

Google Cloud has unveiled its new A3 instances, which will be generally available next month. These instances are powered by NVIDIA’s H100 Tensor Core GPUs, designed to cater to the increasing demands of generative AI and large language models (LLMs). The A3 instances are expected to offer three times the speed and ten times the networking bandwidth compared to Google’s previous A2 supercomputer.

Deepening Collaboration with NVIDIA

The announcement of the A3 instances follows NVIDIA being named Google Cloud’s Generative AI Partner of the Year. This accolade underscores the profound collaboration between the two tech giants in accelerating generative AI on Google Cloud. NVIDIA’s CEO, Jensen Huang, joined Google Cloud’s CEO, Thomas Kurian, at the Google Cloud Next conference to discuss the collaboration.

During their conversation, Huang highlighted the collaboration’s depth, emphasizing the acceleration of the PaxML framework for creating massive Large Language Models using NVIDIA GPU. This Jax-based machine learning framework is designed for training large-scale models, facilitating advanced experimentation and parallelization. Notably, PaxML is now available on the NVIDIA NGC container registry, as mentioned in NVIDIA’s official announcement.

Generative AI Startups Flourishing on Google Cloud

Generative AI startups are increasingly leveraging NVIDIA technology on Google Cloud. For instance, Writer, a startup that uses transformer-based LLMs, enables marketing teams to produce content rapidly. The startup utilizes NVIDIA NeMo, an application framework that aids in curating training datasets, building, and customizing LLMs. Another startup, Runway, employs AI to produce videos in various styles, offering filmmakers and content creators a novel approach to video design.

Google Cloud’s Commitment to AI

Google Cloud’s commitment to AI is evident in its infrastructure enhancements, as highlighted by SiliconANGLE. The company has introduced new tensor processing units (TPU) and announced the upcoming availability of its A3 supercomputer. Mark Lohmeyer, Google Cloud VP and GM of compute and machine learning infrastructure, emphasized Google’s dedication to AI, stating that AI remains a significant aspect of the cloud provider’s workload-optimized infrastructure strategy.

Cloud TPU v5e: A Scalable Solution

Google Cloud has also introduced its cloud TPU v5e, which is touted as the most cost-efficient and accessible cloud TPU to date. These TPUs, or custom ASICs, are designed to accelerate AI and ML workloads. The TPU v5e offers double the training performance per dollar and 2.5 times the inference performance per dollar compared to its predecessor, as reported by SDxCentral.

GKE Enterprise Edition Enhancements

Google Kubernetes Engine (GKE) is another tool that Google Cloud customers utilize for deploying and managing cloud-native AI or ML applications. The hyperscaler has introduced new capabilities for the GKE enterprise edition that support multi-cluster horizontal scaling. These new features are available for cloud GPUs and TPUs, leading to significant improvements in productivity and software deployment times.