What's new in Nsight Perf SDK 2026.3

1.   Added support for GB20B SKU on Windows 11 ARM64 OS and Linux ARM64 OS.

2.   Added support for some GB202 SKUs. See https://developer.nvidia.com/ERR_NVGPU for the detailed SKU info.

3.   Added a terminal-based real-time GPU performance monitoring sample. Please refer to the Monitoring Tool in the Nsight Perf SDK 2026.3 Getting Started Guide for more details.



Deprecations

A period of deprecation has begun for Windows 10, with Windows 11 as the primary focus of testing and support. Nsight Perf SDK may still run on unsupported Windows distributions, however we cannot guarantee compatibility or support outside of the listed distributions.



Limitations

General Limitations


1.    Nsight Perf SDK does not support SLI systems.

2.    Each GPU is limited to one profiling session at a time, across the entire system when in non-MIG mode or (for chips that support it) MIG full device mode due to secure profiler enforcement. When profiling within a MIG, simultaneous profiling sessions are supported (at most, 1 per MIG). In addition, multiple GPUs can be independently profiled concurrently.

3.   NVPW_Device_SetClockSetting() performs a system-global clock change, that is not undone at process exit. The *SetDeviceClockState() family of NvPerfUtility functions have the same behavior, as they are convenience functions around NVPW_Device_SetClockSetting().

4.   Not all counters and metrics are available to the Range Profiler APIs, due to driver limitations. The CounterAvailability APIs exist to reveal which metrics are available.

5.   Quadro GV100 and Titan V are not supported starting from v2022.1.

6.   32 bit X86 is not supported starting from v2022.1.

7.   The Range Profiler API does not support NVLINK-enabled systems, so the HtmlReport examples will not work properly on these systems. Users should disable NVLINK before using related features. However, HudCounters and One-shot Sampling based on the Device Periodic Sampler API are not affected.

8.   prop__latez_input_samples can be negative on Turing and Ampere GPUs when the measured workload has depth test disabled.


D3D12 Profiler


1.   D3D12 Range Profiler is not compatible with the D3D12 debug layer.

2.   D3D12 Range Profiler does not support PushRange and PopRange commands on Bundles (Command Lists created with D3D12_COMMAND_LIST_TYPE_BUNDLE).


Vulkan Profiler


1.   Vulkan Range Profiler is not compatible with Vulkan layers that perform object-wrapping.

2.   Vulkan Range Profiler does not support PushRange and PopRange calls on secondary command buffers (VkCommandBuffer created with VK_COMMAND_BUFFER_LEVEL_SECONDARY.

3.   Vulkan Range Profiler does not support PushRange and PopRange calls on a VkCommandBuffer that is recording with VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT.

4.   vkQueueWaitIdle is required before present when profiling Vulkan applications.


D3D11 Profiler


1.   D3D11 Range Profiler does not support PushRange and PopRange calls on a Deferred Context.