The Extended Brief
NVIDIA Exemplar Cloud: Lessons for Unlocking Full Performance on AI Infrastructure
Brief by The AI News AI newsroom · Jul 30, 2026, 5:12 PM EDT edition
Original reporting by NVIDIA Developer Blog — Elizabeth Goodman · published Jul 30, 2026, 12:00 PM EDT
Misconfiguring kernel settings on new Blackwell or H100 clusters can silently waste up to 12% of your compute throughput, costing millions at scale.
Key points
- Identical NVIDIA H100, GB200 NVL72, or GB300 NVL72 clusters can yield materially different training throughput.
- Partner deployments show 8% to 12% throughput gaps compared to NVIDIA reference architectures.
- These performance gaps occur across identical workloads, models, and global batch sizes.
- Kernel configuration choices frequently cause these performance discrepancies between computing clusters.
From the source
“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.”
“A GB200 NVL72 partner deployment running DeepSeek-V3 Mixture-of-Experts (MoE) FP8 pre-training inside a VM was producing iteration times 12% to 14% longer than the bare-metal RA.”
“Allowing the idle cores to drop to C6 freed power headroom, enabling the busy cores to climb to 3.8 GHz and recover roughly 4% on this workload.”
“A GB300 NVL72 deployment with NVIDIA ConnectX-8 SuperNICs (1.6 Tbps per node) showed a 31% training performance gap on Nemotron-4 15B pre-training.”
“In a virtualized B200 deployment, training throughput was 13%–53% below the NVIDIA reference even though nccl-tests run on the host showed expected performance.”
Practical applications
- Audit kernel configuration on your H100, GB200 NVL72, or GB300 NVL72 clusters against NVIDIA's reference architectures before assuming hardware is the bottleneck.
- Run identical training workloads across your clusters and investigate any throughput gap in the 8-12 percent range as a likely configuration issue.
- Fold a configuration-parity check into cluster acceptance testing so new deployments are validated against reference throughput, not just uptime.
Who should care
Infrastructure and ML platform engineers operating GPU training clusters, since misconfiguration can silently cost 8-12 percent of throughput on hardware they are already paying for.
Context
Large AI training runs are executed on clusters of GPU systems such as NVIDIA's H100 and the newer GB200/GB300 NVL72 rack-scale designs, where small per-node inefficiencies compound across the whole fleet. NVIDIA's Exemplar Cloud findings show that two clusters built from identical hardware, running identical workloads and batch sizes, can differ materially in training throughput. The culprit is frequently kernel configuration choices rather than hardware, with partner deployments showing 8-12 percent gaps versus reference architectures.
What to watch
- Whether NVIDIA publishes concrete reference configurations or tooling that lets operators diagnose these kernel-level gaps automatically.
- Reports from other cluster operators confirming or narrowing the 8-12 percent throughput spread on identical hardware.
Editorial score 3.5 / 5 · significance 4.0 · novelty 3.5 · edge 3.5 · perspective 3.0
Desks: Engineering · Tags: infrastructure, hardware, tooling
Evidence basis: Reviewed from a feed excerpt
This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.