Get Started With Nsight Systems

Download NVIDIA Nsight Systems

Nsight Systems 2026.6.1 is Available Now

Review the supported platforms for NVIDIA Nsight™ Systems to choose the correct version for your host and profiling target.

If profiling from the CLI, pick your platform based on where the CLI will be run. If using the GUI (Full Version) to view reports, do profiling, or do remote profiling, pick your platform based on the host PC architecture where the GUI will be run.

Also review the system requirements before downloading.

Desktop, workstation, and server platforms:

Kubernetes integration:

JupyterLab integration:

Embedded and automotive platforms:

View Nsight Systems documentation.

---

Supported Platforms

Nsight Systems is distributed through multiple packages. Pick a “Profiling Target” column and learn what hosts may be used to profile (local or remote) as well as view reports.

		Profiling Target
		Linux Workstations & Servers	Windows Workstations & Gaming PCs	NVIDIA DPUs & SuperNICs	Jetson & IGX	DRIVE
From Host
	Windows	Remote GUI* Report Viewer**	Local CLI & GUI Remote GUI* Report Viewer**	Remote CLI Remote Report Viewer	Remote Report Viewer	Remote Report Viewer
	Mac	Remote GUI* Report Viewer**	Remote Report Viewer**	Remote Report Viewer	Remote Report Viewer	Remote Report Viewer
	Linux	Local CLI & GUI Remote Report Viewer**	Remote GUI Report Viewer**	Remote CLI Remote Report Viewer	Remote GUI Report Viewer***	Remote GUI Report Viewer***
	DPU / SuperNIC	N/A	N/A	Local CLI	N/A	N/A
	Jetson	N/A	N/A	N/A	Local CLI & GUI Report Viewer*** Remote GUI Report Viewer***	N/A
	DRIVE	N/A	N/A	N/A	N/A	Local CLI

* For x86-64 targets only or opening report collected from a CLI

** Only for reports collected from Windows or Linux PCs & servers of equal or lesser versions

*** Only for reports collected from Jetson or DRIVEOS of equal or lesser versions

---

System Requirements

Nsight Systems is compatible on Windows workstations and PCs, Linux workstations and servers, as well as Jetson and NVIDIA DRIVE Autonomous Machines. Learn about the system requirements and support for your development platform below.

	Windows Workstations and Gaming PCs	Linux Workstations and Servers	Linux Arm Servers	Jetson and Drive Autonomous Machines
Operating Systems	Windows 10 or newer	Ubuntu 26.04, 24.04, 22.04 WSL-Ubuntu 2.0 CentOS 7+* RHEL 8, 9, 10 SLES 15, 16 Debian 12, 13 Fedora 41 KylinOS 11 OpenSUSE 15 Rocky 8, 9, 10	Ubuntu 26.04, 24.04, 22.04 Rhel 8, 9, 10 SLES 15, 16	Jetson Linux DRIVE OS
Target Hardware	GPU: Turing or newer CPU: x86-64 or NVIDIA Arm-based processors	GPU: Turing or newer CPU: x86-64 processors**	GPU: Turing or newer Arm-SBSA servers	NVIDIA IGX, Jetson AGX Orin, Jetson AGX Xavier, Jetson TX2, Jetson TX1, DRIVE AGX Orin, DRIVE AGX Pegasus, DRIVE AGX Xavier, DRIVE PX Parker AutoChauffeur, DRIVE PX Parker AutoCruise
Target Software	64-bit applications only CUDA 12.0+ for CUDA trace Driver 418 or newer***	64-bit applications only CUDA 12.0+ for CUDA trace Driver 418 or newer***	64-bit applications only CUDA 12.0+ for CUDA trace Driver 418 or newer***
Local Profiling	CLI and GUI	CLI and GUI	CLI and GUI	CLI (all platforms), GUI (Jetson Linux only)
Remote Profiling From Platforms	Windows 10+ macOS 13+ Ubuntu 22.04+	Windows 10+ macOS 13+ Ubuntu 22.04+	N/A	Ubuntu 22.04

* For older OS versions, please use Nsight Systems 2020.3
** Intel Haswell architecture or newer is required for LBR sampling backtrace
*** Driver 535 and newer improves GPU profiling stability. Please use the latest driver for the best results. Download here.

---

Release Notes

2026.6.1

Highlights

• Platforms
  • Added Windows-on-Arm GB300 support to match the x86-64 Windows feature set
• Compute
  • CUDA 13.5 support
  • NVTX projection for CUDA graphs extended to include batched memcpy and child graph node scenarios
  • CUDA Kernel tooltips now show CTA/CGA occupancy
  • CUDA synchronization API tooltips now show stream information
• Networking and storage
  • Spectrum-X HFT metrics – configuration file option to set the sampling frequency
  • NCCL GPU operations are shown as one range per operation instead of one range per channel. Communicator rows expand into per-channel rows, using the new support for hierarchical NVTX category names.
  • vClock - improved time alignment accuracy in multi-report view, and added support for recipe time alignment
• AI frameworks
  • NIXL (NVIDIA Inference Xfer Library) KV-cache performance counters
    • Added latency and throughput metrics for analyzing KV-cache efficiency in inference clusters
    • PyTorch profiling – added correlation of backward calls to respective queued GPU workloads, across worker threads and torch.compile transformation
  • Graphics
    • D3D12 DXCGC – added support for the DirectX Compute Graph Compiler
    • Ray tracing ranges - added GPU timestamp ranges around ray dispatches in D3D12 command lists and Vulkan command buffers
    • Trace Vulkan VK_KHR_device_address_commands
  • AI skills & agent support
    • Refactored agentic “nsight-systems“ skill
      • Slimmed down the skill to a lightweight guide for discovering and launching nsys capabilities
      • Merged the agent tools into the nsys CLI and the installed Python environment
      • CLI was expanded to provide a single command-line interface for people and AI agents. See new CLI commands and new recipes below
  • CLI
    • Three new nsys CLI commands:
      • query – run a read-only SQL query on a report or recipe output
      • docs – search the installed documentation
      • survey - extract top-level details of an existing nsys-rep file
    • Capture range callbacks – extended CLI callbacks to --capture-range sessions, let nsys profile and nsys start run user-supplied shell commands when collection starts and before report delivery
    • New --log switch added to collect debug logs
    • New --format=JSONL option for the docs, query, survey and recipe commands.
    • New --show-output=none option for the profile and launch commands. It disables the capture of standard streams. This is the new default on Windows platforms.
  • Analysis and visualization
    • Parquet importer
    • Export improvements
      • Report export was reworked. If an export error occurs, retry with the previous implementation by setting the ``UseLegacyExport=true`` configuration option.
    • CPU Sampling Source info – added support on Windows
    • 3 new recipes
      • gfx_api_analysis - Analyze graphics API durations, percentiles, and call chains
      • gfx_frame_analysis - Analyze CPU frame timing, slow-frame patterns, and related graphics events
      • nccl_pace - Investigate NCCL collective progress and consistency across ranks and iterations
    • Longer explanatory descriptions in nsys recipe --help output
    • Recipe insights
      • 10 of the bundled recipes now generate analysis insights that include observations, suggested further reading, and follow-up analysis steps
      • Insights are included in the recipe output (JupyterLab or JSONL format). A dedicated insights.jsonl file is created to accompany recipe CSV / JupyterLab output
    • CPU events and metrics timeline – improved Y-axis scaling, using a fixed 0–100% range for percentages and a shared maximum across CPUs for other values
  • Bug fixes and stability improvements
    • CPU core event sampling – fixed collection failures on Intel hybrid CPUs with low-power cores, such as Intel Core Ultra 7 265H
    • Recipe dependency installation – fixed failures and improved error messages, including the Python interpreter in the suggested install command
    • CUDA tracing collection via Hardware Event System (HES) on Blackwell and newer GPUs - several bug fixes done
  • NVIDIA Nsight Cloud
    • Nsight Streamer updates
      • Docker
      • Kubernetes
    • Nsight Operator updated
      • Improvements for analysis, OpenTelemetry, Dynamo
      • Learn more via new Docs site
  
  
  

  2026.5.1

  Highlights

  
  
  • NVIDIA Rubin GPU architecture support
    • Performance metrics support for GR100: General Metrics and Compute Triage
  • NVIDIA RTX Spark Support
    • Trace arm64 native and x64 emulated Windows applications
    • CPU core PMU metrics sampling
    • GPU power and temperature metrics
    • NVIDIA Video Codec SDK tracing
  • Compute
    • CUDA 13.4 support
    • CUDA work queue information - Added channel count and channel ID details
    • Multi-process capture ranges - Improved CUPTI buffer flushing when profiling ends through cudaProfilerStop()
    • CUDA kernel details now show if execution model was Tile IR or SIMT
  • Nsight Systems Plugins
    • DCGM - Added “--power-profile” to see if: supported, requested, or enforced
  • Network profiling improvements
    • Nsight Systems can now collect Spectrum-X high frequency performance counters. Counters are read via DTS (DOCA Telemetry Service) and the switch Cumulus OS
  • NVIDIA Nsight Cloud
    • Nsight Streamer updates
      • Docker
      • Kubernetes
    • Nsight Operator updated
      • Improvements for analysis, OpenTelemetry, Dynamo
      • Learn more via new Docs site
  
  
  

  2026.4.1

  Highlights

  
  
  • Graphics improvements
    • D3D12 interfaces support added: ID3D12Device15, ID3D12 ApplicationIdentity, ID3D12 Runtime Validation Control
    • D3D12 Runtime Bypass now supported
    • D3D12 CiG (CUDA-in-Graphics) Streams interception and display
    • Vulkan extensions support added: VK_KHR_maintenance10, VK_EXT_memory_decompression, VK_NV_compute_occupancy_priority, VK_NV_low_latency2
  • CPU metrics improvements
    • Add support for CPU metric sets
  • Agentic AI skill pack preview
    • Focuses on graphics performance analysis and frame stutter analysis
    • Agent-agnostic designed to work with leading AI agents and models
  • CUDA improvements:
    • NVTX projection on “All Streams” hierarchy for a combined view of NVTX without peeking into individual stream hierarchies
    • cuTile name demangling to improve readability
    • Direct3D 12 CUDA-in-Graphics (CiG) stream tracing: CUDA workloads submitted through CiG streams are displayed in the timeline
  • PyTorch improvements:
    • New “--pytorch=function-trace-shapes” option adds function shapes on top of prior option. This is a detail vs overhead tradeoff, so both options are offered
  • Network improvements
    • High-frequency NIC metrics collection via DTS (DOCA Telemetry Service)
    • NCCL recipes improvements, including a new straggler recipe to pinpoint slow ranks
  • Storage improvements
    • S3 (object storage) access summary recipe
    • NVIDIA SCADA metrics
  • Time synchronization
    • Unprivileged virtual clock option, to improve time synchronization across reports to improve alignment when viewing together on a timeline
  • NVTX
    • NVTX Binary Payloads can be exported as dynamic tables with columns per fields
  • Nsight Systems plugins
    • Improve in-process plugins: add initialization stage, add exit callback API, and load plugin libraries into subprocesses
  • NVIDIA Nsight Cloud
    • Nsight Streamer updates
      • Docker
      • Kubernetes
    • Nsight Operator updated
      • Improvements for analysis, OpenTelemetry, Dynamo
      • Learn more via new Docs site
  
  
  

  2026.3.1

  Highlights

  
  
  • Compute
    • CUDA 13.3 support
    • NVTX 3.5.0 support
    • NVTX range projection onto CUDA Graphs improvement
    • PyTorch profiling - forward pass annotations include the original module
  • Graphics
    • Frame latency - new row visualizing combined Reflex SDK stages of the frame. This is the "combined-latency" from initial submission of workloads to display on screen.
    • Resource names appear in the resource allocation event tooltip
  • Networking, storage, and data center
    • High-frequency metrics sampling for NVIDIA Ethernet NICs
    • S3 improvements:
      • Trace AWS C++ SDK support and CRT version 9 SDK
      • Access sum recipe with object and bucket statistics
      • Counters to storage_util_map recipe
    • DCGM collector using the new "--enable=dcgm" command line switch, which collects low frequency metrics about CPUs, GPUs, NICs, NVLink, and NVSwitch
  • CPU & software
    • CPU performance metrics samples:
      • Uncore(chip-level) PMU metrics samples on NVIDIA Grace CPU, integrated into the unified “--cpu-metrics” command line switch
      • LinuxPerf Topdown sampling interval support in the collection script
    • Linux call-stack sampling improvements
      • Source file+line information in the event pane for x64 and Arm
      • “--backtrace” default changed to use frame pointers on x64 for improved depth & analysis, matching Arm
    • GUI support for remote profiling symbol resolution, including remote symbol locations, server addresses, and symbol cache configuration
  • NVIDIA Nsight Cloud
    • Streamer updated to view reports loaded on remote headless servers
      • Docker
      • Kubernetes
    • Operator updated
      • Improvements for analysis, OpenTelemetry, Dynamo
      • Learn more via new Docs site
    • Ubuntu 26.04 support

---

Feature Table

Feature	Linux Workstations and Servers	Windows Workstations and Gaming PCs	Jetson Autonomous Machines	DRIVE Autonomous Vehicles
View system-wide application behavior across CPUs and GPUs
CPU cores utilization, process, & thread activities	yes	yes	yes	yes
CPU thread periodic sampling backtraces	yes*	yes	yes	yes
CPU thread blocked state backtraces	yes**	yes	yes	yes
CPU performance metrics	yes	no	yes	yes
GPU workload trace	yes	yes	yes	yes
GPU context switch trace	yes	yes	yes	yes
SOC hypervisor trace	-	-	-	yes
SOC memory bandwidth sampling	-	-	yes	yes
SOC Accelerators trace	-	-	Xavier+	Xavier+
OS Event Trace	ftrace	ETW	ftrace	ftrace, QNX kernel events
Investigate CPU-GPU interactions and bubbles
User annotations API trace NVIDIA Tools Extension API (NVTX)	yes	yes	yes	yes
CUDA API	yes	yes	yes	yes
CUDA libraries trace (cuBLAS, cuDNN & TensorRT)	yes	no	yes	yes
OpenGL API trace	yes	yes	yes	yes
Vulkan API trace	yes	yes	no	no
Direct3D12, Direct3D11, DXR, & PIX APIs	-	yes	-	-
OpenXR	-	yes	-	-
OptiX	7.1+	7.1+	-	-
Bidirectional correlation of API and GPU workload	yes	yes	yes	yes
Identify GPU idle and sparse usage	yes	yes	yes	yes
Multi-GPU Graphics trace	OpenGL and Vulkan	Direct3D12, OpenGL and Vulkan	-	-
Trace graphics resource migration between VRAM and System Memory	-	yes	-	-
Ready for big data
Fast GUI capable of visualizing in excess of 10 million events on laptops	yes	yes	yes	yes
Additional command line collection tool	yes	no	no	no
NV-Docker container support	yes	-	-	-
NVIDIA GPU Cloud support	yes	-	-	-
Minimum user privilege level	user	administrator	root	root

* On Intel Haswell and newer CPU architecture
** Only with OS runtime trace enabled. Some syscalls such as handcrafted assembly may be missed. Backtraces may only appear if time threasholds are exceeded.

---

Archives

Access older versions of Nsight Systems in the Gameworks Download Center.
View older version release notes in the Nsight System’s documentation archive.

---

Resources

Nsight Systems Documentation

You can also learn about installing & using the NVIDIA Tools Extension API (NVTX) here.

Nsight Tools Tutorial Center

Access the latest resources to get started with Nsight Systems.

---

Access Self-Paced Training

Nsight Systems Documentation

Get hands on training for Nsight Systems with self-paced online courses from the NVIDIA Deep Learning Institute.

See more courses on Accelerated Computing for Developers.

---

Tutorial Sessions

---

Video Series

Learn about using Nsight Systems for CUDA Development in the CUDA Developer Tools tutorial series.

---

Support

To provide feedback, request additional features, or report support issues, please use the Developer Forums .