NVIDIA Exemplar Cloud: Why Identical AI Clusters Lose Up to 12% of Performance

NVIDIA has published a breakdown of its Exemplar Cloud program in the Developer Blog: two AI clusters built on identical H100, GB200 NVL72 or GB300 NVL72 accelerators routinely differ in training throughput by 8% to 12%, on the same task, with the same model and the same global batch size.

Why identical clusters deliver different results

The cause of the gap is not the hardware but configuration choices. According to the NVIDIA Developer Blog, the 8% to 12% discrepancy between partner deployments and the company's Reference Architecture (RA) stems from a stack of configuration decisions at the operating system kernel level. Two clouds with identical H100 or GB200 NVL72 GPUs can show fundamentally different training throughput if their software layers are set up differently from what NVIDIA recommends.

To put the scale in perspective: on a cluster of a thousand accelerators, a 10% gap amounts to dozens of hours of compute time every week. At market GPU rental prices, that translates into hundreds of thousands of dollars of lost performance per year.

• Systems compared: NVIDIA H100, GB200 NVL72, GB300 NVL72
• Documented gap: 8% to 12% in training throughput
• Conditions: same task, same model, same global batch size
• Source of the gap: a "stack of configuration decisions at the kernel level," according to the NVIDIA Developer Blog

What is NVIDIA Exemplar Cloud

Exemplar Cloud is an NVIDIA partner program designed to help cloud providers reach the performance of the reference architecture. The company analyzes partners' specific deployments, compares them against the RA and identifies configuration bottlenecks that quietly erode cluster efficiency. Typical areas for optimization include network stack parameters, NVLink and InfiniBand settings, Linux kernel configuration and NUMA topology.

The program is especially relevant right now: GB200 NVL72 and GB300 NVL72 are NVIDIA's newest racks, built on the Blackwell architecture, and they are being deployed at scale by the world's largest providers in 2025–2026. Competition for AI customers is fought on real-world throughput, and a 10% gap between two identically equipped clouds can decide the outcome of a major tender.

"We routinely see a discrepancy of 8% to 12% between partner installations and our reference architecture on the same task, the same model and the same global batch size," the NVIDIA

Developer Blog post states.

What this means

For ML engineers and DevOps teams running GPU clusters, the key takeaway is simple: the right hardware is necessary but not sufficient. The 8–12% of performance lost to configuration is real money and real model training time. Exemplar Cloud offers a systematic approach: compare your deployment against NVIDIA's reference and fix the specific bottlenecks instead of guessing why throughput falls short of the theoretical maximum. For cloud providers, matching RA performance becomes both a competitive advantage and an operational saving.

Frequently asked questions

What is the NVIDIA Reference Architecture?

The NVIDIA Reference Architecture (RA) is a reference configuration of the hardware and software stack that NVIDIA publishes for H100, GB200 NVL72 and GB300 NVL72 systems. It serves as the baseline: a partner deployment's performance is compared against the RA to pinpoint specific discrepancies.

Why is an 8–12% gap critical for cloud AI?

On a cluster of hundreds or thousands of GB200 NVL72 accelerators, a 10% gap in training throughput means an equivalent 10% in excess compute costs, or the same share of lost GPU rental revenue compared with a competitor running identical systems with a better-tuned stack.