NVIDIA nvperf-core: The Open Headers Behind Always-On GPU Profiling for AI

NVIDIA has published nvperf-core, a new GitHub repository with the C headers for the NvPerf Core API behind its upcoming Nsight Perf SDK for AI. The repository was created on November 2, 2026, and its first release, 2026.4.1, followed minutes later. The headers are Apache-2.0. The runtime libraries are proprietary. The product they feed, an always-on profiling layer for long-running AI jobs, is still labeled "Coming Soon" on NVIDIA's own page.

That makes this a quiet but telling release. It is not a model or a benchmark. It shows where NVIDIA wants GPU observability to go for training and inference at scale, and it shows how much of that stack will stay closed.

TL;DR: nvperf-core at a glance

What does the repository actually contain?

Very little code, by design. The README says the repository "provides the C headers necessary to interface with the NvPerf Core API which powers the NVIDIA Nsight Perf SDK for AI." The include folder holds separate headers for each target API, with a common header, host and target headers per API, and a versions header. We saw files for CUDA-adjacent device access, OpenGL, EGL, Vulkan including Vulkan SC, and Direct3D 11 and 12. There are Linux folders for desktop x64 and aarch64.

The README is blunt about limits: the repository "does not contain any setup instructions or sample applications," and it is "currently not accepting contributions." Prebuilt runtime libraries are published on the releases page with a SHA256SUMS file you can verify with sha256sum --check.

The one release, tagged v2026.4.1, is titled "NvPerf Core 2026.4.1" with the note "Initial release, version 2026.4.1." Its assets are two Linux archives, one for aarch64 and one for x86_64. The GitHub API reported the repository with the C language tag, no stars and no homepage.

What is the Nsight Perf SDK for AI?

Here the primary source is NVIDIA's Nsight Perf SDK page. It describes two workflows. The first, graphics profiling, is the long-standing toolbox for DirectX, Vulkan and OpenGL that collects GPU performance metrics directly from an application. The second, "Always-On AI Profiling," is the new one and is marked Coming Soon.

NVIDIA frames the problem plainly. "Traditional profiling works when a workload can be paused or replayed and an issue can be reproduced while a detailed profiler is attached." Long-running, distributed training and inference jobs make that hard: "Intermittent behavior may disappear, and recreating the full environment can be costly or impractical."

The proposed answer is "an always-on NVIDIA CUDA/GPU collection layer that preserves selected performance metrics and execution context from the original run." NVIDIA stresses that it is "a focused record, not a full system trace or kernel profile." The pitch has three parts:

• Preserve the evidence. Run the collection layer next to the workload so metrics from an intermittent event are available afterward.
• Stay alongside the workload. Avoid the overhead and data volume of conventional full-detail capture.
• Narrow the next investigation. Use the record to pick the time interval and activity worth a deeper look in Nsight Systems or Nsight Compute.

For analysis, NVIDIA says Google Perfetto "is the supported analysis and visualization experience available today" for data collected through the SDK. You inspect selected metrics and execution context on a timeline around the interval of interest.

Where does it sit next to DCGM and OpenTelemetry?

NVIDIA positions the SDK between monitoring and deep profiling. Monitoring systems such as NVIDIA DCGM and OpenTelemetry "can signal that a system may be unhealthy," while the SDK "preserves performance evidence tied to the workload activity that produced it." Nsight Systems then provides system-wide tracing across CPUs, GPUs, operating-system activity and libraries, and Nsight Compute provides detailed per-kernel metrics.

If you run GPU clusters, this is a familiar gap. Dashboards tell you a node dipped. They rarely tell you which collective, kernel or data-loading step caused it. For a view of what DCGM-level telemetry looks like in a real deployment, see our piece on a 32-GPU B300 cluster and its DCGM exporter.

Where does the NvPerf name come from?

The "NvPerf" prefix is old. NVIDIA's developer site describes the earlier PerfWorks API as replaced by an updated C++ API under the Nsight Perf SDK name, and the SDK's graphics edition has long let developers pull GPU counters straight from their own code. What is new here is the packaging: a public repository for the core headers, a tagged release with checksums, and a stated link to an AI-focused product. In other words, NVIDIA is taking a graphics-era counter library and aiming it at training and inference clusters. Whether the same counters give useful signals for distributed CUDA workloads is the open question the Coming Soon product has to answer.

Why would anyone care about a headers-only repo?

Three reasons.

First, it signals a product direction. Large training runs now cost enough that a mysterious slowdown is expensive. A tool built to catch intermittent slowdowns in the original run, instead of asking you to reproduce them, addresses a real operational pain. NVIDIA has been working on adjacent AI-for-chips problems too, as in our coverage of Build-a-Chip Bench and NVIDIA's ChipAgents.

Second, it shows a hybrid license model. Open headers let third-party tools compile against the API. Closed runtime libraries keep the implementation under NVIDIA's terms. The repository's LICENSING.md says the headers carry an SPDX-License-Identifier: Apache-2.0 notice, while the libnvperf_grfx_host.so runtime libraries "are proprietary NVIDIA software governed by the NVIDIA Nsight Perf SDK License (28Sept2022)" and each archive includes a copy of that EULA. Anyone planning to embed the library in a product should read that license first.

Third, it touches security-sensitive ground. A collection layer that sits beside production workloads and records GPU state is powerful. Our post on the Eclypsium findings about NVIDIA AI hardware tampering shows why teams treat anything with low-level device access carefully. NVIDIA has not described, in the sources we read, what data the collection layer stores or how it is protected.

What can a developer do right now?

Honestly, not much beyond reading. The README sends you to the Nsight Perf SDK for AI repository for setup. When we requested that repository's GitHub API endpoint and README, GitHub returned 404, and a web search found no public repository, only NVIDIA's Coming Soon page. That could mean the repo is not public yet. We do not know the release date, and NVIDIA has not given one in the sources we found.

What you can do today:

1. Read the headers. Verify the API surface and which target APIs NVIDIA exposes.
2. Use the existing graphics SDK if you do graphics profiling. NVIDIA's page describes a GPU Periodic Sampler and a real-time performance HUD there.
3. Plan for Perfetto. If you want to use the AI workflow when it ships, the timeline analysis target is Google Perfetto, according to NVIDIA.
4. Keep your current profilers. Nsight Systems and Nsight Compute are still the tools for deep dives, and NVIDIA says the SDK complements them.

For teams choosing hardware or workstations to run such tools, our look at the Gigabyte W775 GB300 deskside workstation versus DGX Spark covers one end of NVIDIA's lineup, and the NVIDIA long-context world action model paper shows the kind of heavy job that makes profiling worth the effort.

What is claimed, and what is verified?

Note the missing piece: NVIDIA publishes no overhead figure, no supported-workload list and no sample trace in the sources we reviewed. "Without the overhead and data volume of conventional full-detail capture" is a design goal, not a measured result. Treat any performance claim as unverified until independent users report numbers.

The bottom line

nvperf-core is a small, early release with a clear purpose: give developers a stable C interface to NVIDIA's performance-counter machinery ahead of an always-on profiling product for AI jobs. The open headers help tool builders. The closed runtime keeps NVIDIA in control. If you operate large training or inference clusters, put the Nsight Perf SDK for AI on your watch list, and judge it on overhead numbers, supported frameworks, and data-handling details once they appear.

Repository details reflect GitHub and NVIDIA pages as read on November 3, 2026 and may change.

Related reading

• Lava 32-GPU B300 cluster and the DCGM exporter
• Build-a-Chip Bench and NVIDIA ChipAgents
• Eclypsium: NVIDIA AI hardware tampering, what is confirmed
• Gigabyte W775 GB300 deskside workstation vs DGX Spark
• NVIDIA long-context world action model
• Official: nvperf-core on GitHub, NVIDIA Nsight Perf SDK