NVIDIA and Microsoft Expand Accelerated Computing From Azure to Windows PCs
NVIDIA and Microsoft are extending their full‑stack partnership to deliver Azure virtual machines powered by NVIDIA Vera Rubin NVL72 systems and to expose NVIDIA open models through Microsoft Foundry, while also adding edge AI options. Azure’s Vera Rubin preview achieved MLPerf Training v6.1 scale on 10,240 GPUs (GB200 NVL72) for DeepSeek‑V3 and led Inference v6.1 on DeepSeek‑R1 with >10% higher interactive throughput. Enterprises can now customize Nemotron 3.5 models via Foundry’s Programmed Training, as EY demonstrated a 55% accuracy lift and 15× cost reduction versus proprietary models. These advances give engineers a path to high‑scale, power‑efficient, and confidential AI workloads from cloud to Windows PCs.
⚡ Key Takeaways
- Azure scaled to **10,240 NVIDIA GPUs** (GB200 NVL72) in MLPerf Training v6.1 on the DeepSeek‑V3 model.
- Azure’s **NVIDIA Vera Rubin NVL72** systems are the core compute platform for demanding reasoning, agentic apps, and large‑scale training/inference.
- Power‑efficiency is addressed through **NVIDIA MaxLPS**, targeting higher performance per watt in data centers.
- Enterprises can customize open models using **Microsoft Foundry Programmed Training**, which managed the underlying infrastructure while allowing post‑training workflow control.
- Confidential computing is enabled via **NVIDIA Confidential Computing and attestation**, requiring verification of the compute environment before releasing sensitive assets.
Engineers can now provision massive, power‑efficient GPU clusters on Azure for both training and inference, while leveraging Managed Compute to fine‑tune open‑source Nemotron models with proven cost and accuracy gains, all under hardware‑rooted security guarantees.
✅ Practical Steps
- Deploy Azure VMs with **NVIDIA Vera Rubin NVL72** systems for large‑scale training or inference workloads.
- Use **Microsoft Foundry Programmed Training** to adapt NVIDIA Nemotron models to your domain data, following the same workflow EY used for corporate purchasing record classification.
- Deploy the customized model weights via **Foundry Managed Compute** to connect model customization directly to production services.
Want the full story? Read the original article.
Read on NVIDIA Blog ↗