Maximize GPU efficiency
Introducing Continuous GPU Profiling. See the last hours of performance insights.
Trusted by
Optimize GPU performance by profiling GPU workloads
Correlate GPU hardware metrics with CPU activity in a single view to find exactly where cycles are being wasted.
Optimize GPU workloads precisely
See GPU utilization in detail and optimize with precision
NVIDIA CUDA
Supports CUDA 11+ and modern NVIDIA GPUs, including instruction-level PC sampling with stall-reason attribution
Python
Supports multiprocessing and top frameworks
PyTorch
Understand GPU usage across ML pipelines
Intel upon request
Compatible with oneAPI and Intel Extension for PyTorch
AMD upon request
Visibility into HIP kernels and AMD Instinct accelerators
Frequently asked questions
With Continuous Profiling for GPUs you can understand GPU usage in production via multiple metrics, correlate CPU activity with GPU utilization, and visualize CPU time spent with flame charts to find out what's keeping the GPU waiting.
No, you don't have to use Kubernetes or even containers. The agent needs Linux (x86 or ARM) and the GPU Metrics Agent needs Linux and NVIDIA drivers.
Right now, the GPU metrics agent collects the following metrics (more can be added).
- GPU Utilization
- GPU Memory Utilization
- GPU Power Usage
The Continuous Profiler for GPUs supports profiling GPU workloads running on NVIDIA GPUs. If you are interested in us adding more accelerators, such as AMD, Intel, Apple GPUs, or TPUs, please reach out to us.
PC sampling captures not just which GPU instruction was running, but why it stalled (e.g. a memory dependency or a synchronization wait) at very high sample rates. It’s traditionally been limited to interactive profilers because of overhead.
Boost your apps, maximize efficiency, and lower expenses
Every idle GPU cycle is money spent for nothing. See exactly where utilization drops, correlate it with what the CPU was doing at that moment and fix the gap