NVIDIA and Oracle announced a collaboration to accelerate AI and data processing for enterprises at the Oracle CloudWorld conference. Oracle Cloud Infrastructure (OCI) launched the first zettascale OCI Supercluster, powered by over 100,000 of NVIDIA’s latest GPUs. These systems can scale up to 131,072 Blackwell GPUs, delivering 2.4 zettaflops of peak AI compute to the cloud.
Oracle previewed liquid-cooled bare-metal instances to enhance training workloads and deliver real-time inference at scale. These instances will connect up to 65,536 H200 GPUs with NVIDIA ConnectX-7 NICs over RoCEv2 cluster networking. Companies like Reka, a foundation model startup, are using these clusters to develop advanced multimodal AI models for enterprise agents.
Dani Yogatama, cofounder and CEO of Reka, said, “Reka’s multimodal AI models, built with OCI and NVIDIA technology, empower next-generation enterprise agents to comprehend and interact with the world.”
Oracle Autonomous Database is gaining NVIDIA GPU support to accelerate data processing workloads. NVIDIA and Oracle demonstrated three capabilities showing how NVIDIA’s accelerated computing platform can enhance components of generative AI pipelines: bulk vector embeddings, vector graph index generation, and NVIDIA microservices. NVIDIA and Oracle are collaborating to deliver infrastructure worldwide, addressing data residency needs.
Brazil-based startup Wide Labs trained and deployed Amazonia IA, a large language model for Brazilian Portuguese, within OCI’s Brazilian data centers, ensuring data sovereignty. Enterprises can accelerate task automation on OCI by deploying NVIDIA software such as NIM microservices with OCI’s scalable cloud solutions. This enables quick adoption of generative AI and the creation of agentic workflows for complex tasks.
Oracle announced that it is now taking orders for a groundbreaking 2.4 zettaFLOPS cluster, boasting “three times as many GPUs as the Frontier supercomputer.” However, the claims come with a significant caveat concerning precision in measuring performance. The 131,072 Blackwell accelerators that constitute Oracle’s new cluster are capable of producing 2.4 zettaFLOPS of sparse FP4 compute. This is primarily a marketing figure, as it uses 4-bit floating points that are common in AI applications but not in the traditional scientific computing benchmarks that typically use 64-bit double precision.
Oracle’s claimed 2.4 zettaFLOPS performance is impressive but somewhat misleading.
HeatWave capabilities enhance generative AI
When measured using more common AI precisions like 16-bit floating points, the cluster’s performance drops to about 459 exaFLOPS.
While this number is still substantial, it lacks the eye-catching appeal of “zettascale.”
The potential of Oracle’s Blackwell supercluster remains high. At 64-bit precision, its peak performance ranges between 5.2 and 5.9 exaFLOPS, which is more than three times that of the AMD Frontier system. However, scaling to peak performance is often restricted by interconnect limitations.
Oracle also plans to release Blackwell-based superclusters, including Nvidia’s flagship GB200 NVL72 rack-scale systems, starting in the first half of 2025. Current offerings include H100 and H200 superclusters that can scale to 16,384 and 65,536 GPUs, respectively. Oracle announced new HeatWave capabilities, including innovations to help organizations more easily and securely take advantage of generative AI both in Oracle Cloud Infrastructure (OCI) and in Amazon Web Services (AWS).
Additional capabilities help customers implement lakehouse and machine learning applications for various use cases, as well as improve the performance and manageability of transactional applications. Edward Screven, chief corporate architect, Oracle, said, “Our track record of delivering powerful new HeatWave capabilities that are automated, integrated, and secure now extends to AI. Some organizations, such as SmarterD, are building new generative AI applications on HeatWave and moving into production in less than a month.
To enable more customers to take advantage of our innovations, including HeatWave GenAI with in-database LLMs, we are making them available natively on AWS.”
Hans Ospina, CTO and founder, SOCOBOX, said, “The integration of generative AI in HeatWave is a major leap forward for us. By bringing in-database LLMs, automated vector processing, AutoML, and lakehouse into our workflows, we can now deliver powerful AI-driven insights and applications without the overhead of external tools.”
New HeatWave capabilities on AWS, also available on OCI, include HeatWave GenAI, HeatWave Lakehouse, native JavaScript support, and HeatWave Autopilot indexing. HeatWave GenAI integrates automated, secure generative AI, enabling developers to build new AI applications without needing AI expertise.
HeatWave MySQL enhances OLTP workloads by leveraging MySQL Database Enterprise Edition features with new capabilities like hypergraph optimizer, integration with OCI Ops Insights, and bulk ingest. HeatWave Lakehouse queries vast data stores with high performance, with new features such as writing results to object storage and automatic change propagation. HeatWave AutoML supports building, training, and explaining machine learning models at no additional cost, with new capabilities such as storing and processing larger models, topic modeling, data drift detection, and semi-supervised log anomaly detection.
HeatWave is now available in the OCI Always Free Service, allowing organizations to develop and run small-scale applications using HeatWave MySQL, analytics, machine learning, JavaScript, and process data in object storage with up to 50 GB of storage and backup for an unlimited time, alongside $300 of credits for other OCI services for up to 30 days.
