Oracle Cloud Infrastructure (OCI) has achieved NVIDIA Exemplar Cloud validation for NVIDIA GB300 NVL72 and NVIDIA HGX B300 platforms, further demonstrating its ability to deliver consistent, reproducible AI performance on NVIDIA Blackwell infrastructure at scale.
This milestone builds on OCI’s earlier validation for NVIDIA GB200 NVL72, extending proven performance across the latest generation of Blackwell GPUs for AI training.
Why this matters
As AI models continue to grow in size and complexity, organizations need more than peak benchmark numbers—they need infrastructure that delivers predictable performance across large, distributed workloads.
NVIDIA Exemplar Cloud provides an independent benchmark for evaluating cloud infrastructure. To earn the designation, cloud providers must demonstrate that NVIDIA reference workloads run within 5% of NVIDIA’s target performance. OCI’s successful validation on both GB300 NVL72 and HGX B300 platforms confirms that customers can expect consistent, production-ready performance for demanding AI workloads.
This combination of high-performance compute and high-bandwidth networking enables customers to scale training and inference workloads with confidence while maintaining reproducible results across large GPU deployments.
Engineered for large-scale AI
OCI’s BM.GPU.GB300.4 shape is designed to maximize both scale-up and scale-out performance. Each node provides:
- 4 NVIDIA Blackwell Ultra GPUs
- 144 OCPUs
- 960 GB host memory
- 400 Gbps networking bandwidth
Eighteen nodes combine to form a 72-GPU GB300 NVL72 rack-scale domain. Within each rack, NVIDIA NVLink connects all 72 Blackwell Ultra GPUs into a unified scale-up domain, while multi-rail NVIDIA Spectrum-X Ethernet networking enables efficient scale-out training across OCI AI clusters.
For NVIDIA HGX B300 deployments, OCI uses the BM.GPU.B300.8 shape. This platform features NVIDIA B300 GPUs and an optimized eight-rail Spectrum-X Ethernet configuration designed for high-performance distributed AI workloads.
Validated across industry-leading AI models
The validation included a diverse set of representative large language model training workloads, including:
- DeepSeek V3
- GPT-OSS 120B
- Kimi-K2
- Llama 3.1
- Nemotron-H
- Nemotron 3
- Qwen3
Across this workload suite, both OCI GB300 NVL72 and HGX B300 platforms successfully met the NVIDIA Exemplar Cloud performance threshold, demonstrating reliable performance across modern foundation model architectures.
Built for production-scale AI
Achieving Exemplar Cloud validation reflects the engineering investments behind OCI’s NVIDIA Blackwell infrastructure, from compute and networking to systems software and cloud operations.
Developed in close collaboration with NVIDIA, OCI’s Blackwell deployments are designed to help customers move beyond experimentation and confidently deploy AI workloads at production scale.
For customers, the outcome is straightforward: The same infrastructure validated through these benchmarks is designed to power real-world AI training environments requiring thousands of GPUs, high-throughput networking, advanced precision formats, and consistent performance throughout long-running distributed jobs.
Setting a new standard for cloud AI
OCI believes customers should be able to trust that published benchmark results reflect the performance they can achieve in production.
With NVIDIA Exemplar Cloud validation now spanning HGX B200, GB200 NVL72, HGX B300, and GB300 NVL72 platforms, OCI continues to demonstrate reproducible, benchmark-validated performance across the NVIDIA Blackwell portfolio.
Whether customers require node-level GPU instances or rack-scale NVL systems, OCI provides the high-performance, scalable infrastructure needed to accelerate the next generation of AI innovation with confidence.
