NVIDIA Exemplar Cloud: Lessons for Unlocking Full Performance on AI Infrastructure

<img width="768" height="431" src="https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-768x431.png" class="webfeedsFeaturedVisual wp-post-image" alt="" style="display: block; margin-bottom: 5px; clear:both;max-width: 100%;" link_thumbnail="" decoding="async" srcset="https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-768x431.png 768w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-179x100.png 179w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-300x168.png 300w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-625x350.png 625w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-645x362.png 645w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-660x370.png 660w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-500x280.png 500w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-160x90.png 160w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-362x203.png 362w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-196x110.png 196w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-960x538.png 960w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6.webp 1024w" sizes="(max-width: 768px) 100vw, 768px" title="image6" />Two AI computing clusters built from identical NVIDIA H100, GB200 NVL72, or GB300 NVL72 systems can deliver materially different training throughput. We...<img width="768" height="431" src="https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-768x431.png" class="webfeedsFeaturedVisual wp-post-image" alt="" style="display: block; margin-bottom: 5px; clear:both;max-width: 100%;" link_thumbnail="" decoding="async" loading="lazy" srcset="https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-768x431.png 768w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-179x100.png 179w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-300x168.png 300w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-625x350.png 625w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-645x362.png 645w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-660x370.png 660w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-500x280.png 500w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-160x90.png 160w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-362x203.png 362w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-196x110.png 196w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6-960x538.png 960w, https://developer-blogs.nvidia.com/wp-content/uploads/2026/07/image6-6.webp 1024w" sizes="auto, (max-width: 768px) 100vw, 768px" title="image6" /><p>Two AI computing clusters built from identical NVIDIA H100, GB200 NVL72, or GB300 NVL72 systems can deliver materially different training throughput. We routinely see 8% to 12% gaps between partner deployments and the corresponding NVIDIA reference architecture (RA) on the same workload, same model, same global batch size. The cause is often a stack of configuration choices in the kernel…</p> <p><a href="https://developer.nvidia.com/blog/nvidia-exemplar-cloud-lessons-for-unlocking-full-performance-on-ai-infrastructure/" rel="nofollow" data-wpel-link="internal" target="_self">Source</a></p>