Get Real-Time Visibility into GPU Usage Across Kubernetes Clusters | NVIDIA Technical Blog
… For any environment beyond a local development cluster, update credentials through values.yaml before exposing the dashboard to wider teams. …
Search
Nsight Systems - Get Started
… Export changes Text and JSON export options are removed from both the CLI and the GUI. Using these options e.g. nsys export --type=text ... or nsys profile --export=json ... will result in an error. …
NVIDIA CUDA 13.3 Enhances GPU Development with Tile Programming in C++, Compiler Autotuning, and Python Updates | NVIDIA Technical Blog
… NUMA-aware pools from cuda.core import DeviceMemoryResource, PinnedMemoryResource, PinnedMemoryResourceOptions, ManagedMemoryResource, ManagedMemoryResourceOptions NUMA-pinned host memory pinned = PinnedMemoryResource PinnedMemoryResourceOptions numa id=0 CUDA graphs: stream capture and explicit co… …
Accelerating Vision AI Pipelines with Batch Mode VC-6 and NVIDIA Nsight | NVIDIA Technical Blog
… These stalls point to shared memory loads LDS in Figure 6 , as dynamic indexing into a local array would otherwise result in slow local memory access. …
Running Large-Scale GPU Workloads on Kubernetes with Slurm | NVIDIA Technical Blog
… To enable this, add the following to the NVIDIA GPU Operator Helm values: dcgmExporter: hpcJobMapping: enabled: true directory: /var/lib/dcgm-exporter/job-mapping Worker pods mount the same directory, allowing Slurm prolog/epilog scripts which run at job start and completion to write job mapping fi… …
Accelerating Data Processing with NVIDIA Multi-Instance GPU and Locality Domains | NVIDIA Technical Blog
… Expect additional publications as updated capabilities and practices become available. T he term ‘Locality Domain’ is used in the post to describe GPU-internal memory locality exposed through MIG. …
Making Softmax More Efficient with NVIDIA Blackwell Ultra | NVIDIA Technical Blog
… The following kernel code isolates the exponential instructions to measure the raw cycle count without interference from global memory latency or other arithmetic operations. …
Tuning Flash Attention for Peak Performance in NVIDIA CUDA Tile | NVIDIA Technical Blog
… The online softmax algorithm We maintain two running values for each row: \ m i\ : The maximum value seen so far for numerical stability \ l i\ : The sum of exponentials seen so far the softmax denominator When we process a new tile with values \ x {new}\ : Update the maximum : \ m {new} = \max m i… …
Vulkan Driver Support
… Possible workarounds: Disable flipping in nvidia-settings uncheck "Allow Flipping" in the "OpenGL Settings" panel Disable UBB run 'nvidia-xconfig --no-ubb' Use a composited desktop Bug fixes NVX multiview per view attributes and geometry passthrough shaders Fix subpass dstSubpass=VK SUBPASS EXTERNA… …
MDL SDK
… For a complete list of changes and fixes please see the release notes shipped with the SDK Archive MDL SDK 2025.0.3 Windows / Linux / MAC Enhancements: Updated recommended MaterialX version to 1.39.4 Various example improvements including MaterialX support for the df vulkan example Improved databas… …