CoreWeave named first NVIDIA Exemplar Cloud for GB200 NVL72 training

What the Exemplar Cloud status covers

The NVIDIA Exemplar Cloud initiative was launched to provide rigorous, standardized benchmarking across cloud platforms, with the aim of ensuring transparency and reproducibility. CoreWeave says it worked with NVIDIA on the GB200 NVL72 tests and that its infrastructure, optimized by CoreWeave Mission Control, consistently exceeded target performance across critical test cases.

This is not the company's first such designation. Earlier this year, CoreWeave became an NVIDIA Exemplar Cloud for the NVIDIA H100 GPU, using up to 1,024 GPUs and a 30 billion parameter Llama 2-style model.

The certification process, as described by CoreWeave, tests performance and stability under extreme load. It focuses on three areas: cluster goodput and efficiency, performance scaling across large distributed GB200 NVL72 clusters, and hardware and system resilience during prolonged training sessions.

Benchmark results on 512 GPUs

Performance was measured using MFU, which CoreWeave calls the independent gold standard for measuring efficiency in large model training. All four results below were achieved on 512 GPUs and compared with the NVIDIA reference target:

• DeepSeek v3 (BF16): 1.9% greater MFU
• Grok-1 314B (BF16): 4.7% greater MFU
• Llama 3.1 405B (FP8): 3.8% greater MFU
• Llama 3 70B (FP8): 2.4% greater MFU

The margins are modest in percentage terms, but they show the cluster running above the targets NVIDIA set rather than merely meeting them.

How the test cluster was run

CoreWeave stresses that it did not use a "cherry-picked" performance lab environment. The benchmarks ran on a standard CoreWeave cloud cluster made up of 8 racks of NVIDIA GB200 NVL72 systems, connected by NVIDIA Quantum-2 InfiniBand networking.

According to the company, CoreWeave Mission Control uses a large dataset gathered from operating hundreds of thousands of NVIDIA GPUs to detect and mitigate failures predictively instead of relying on passive monitoring. It continuously tests components and replaces suboptimal ones before they affect a customer's workload.

The software stack described in the announcement includes:

• CoreWeave Bare Metal servers: direct access to GPU resources without hypervisors.
• CoreWeave Kubernetes Service (CKS): the base runtime, integrated with Mission Control.
• CoreWeave Slurm on Kubernetes (SUNK): topology-aware scheduling, which CoreWeave says allows maintenance and predictive fault management without interrupting active jobs.

Monitoring used CoreWeave Cabinet Visualizer for rack health, Mission Control observability for infrastructure, and Weights & Biases for job-level metrics.

Straggler detection and automatic recovery

To keep performance uniform across all 8 racks, CoreWeave used a recently launched Mission Control feature called GPU Straggler Detection. It provides real-time collective metrics, including Bus Bandwidth, and automatically detects hardware lockups and pinpoints the root cause. In one displayed iteration, an underperforming component was intentionally added to the cluster to demonstrate the detection.

CoreWeave says the benchmark completed with no actions by cluster administrators to isolate faults or underperforming components. When potential hardware faults appeared, Mission Control isolated the issue, surfaced the data in the SUNK and Weights & Biases dashboards, and triggered an automatic job restart on healthy nodes.

What it means for customers

For teams training large models, the practical claim is predictability. CoreWeave argues that published, standardized benchmark data lets developers, researchers and engineers plan deployments with confidence and hold providers accountable for performance. The company says the results reflect what customers consistently receive across hardware and software generations, and that they help reduce training time and cost. These are the company's own claims, based on tests run with NVIDIA's reference targets.