Zymtrace

What Zymtrace is

Zymtrace profiles CPU and GPU workloads to locate performance bottlenecks in AI training and inference. Its CPU profiler supports languages including Python, Java, Go, Rust, C/C++, Node.js, Ruby, .NET, PHP, Perl, Zig, and Scala. GPU profiling correlates CPU and GPU execution and analyzes CUDA kernels, GPU instructions, and stall reasons. Monitoring includes GPU utilization, memory, power, temperature, SM efficiency and occupancy, Tensor Core utilization, and data transfer rates. Listed inference engines include vLLM, Ollama, llama.cpp, and SGLang. Zymtrace is currently limited to Linux machines because it relies heavily on eBPF; Windows is not supported. Plans are self-hosted, with the maker stating that data stays within the customer's infrastructure. The air-gapped option performs processing and analysis locally. Each instance has a REST API, and profiling data can be exposed through MCP for AI agents. The product supports OpenTelemetry resource attributes and export in the standard OpenTelemetry format. Basic is free forever and includes CPU profiling for up to five concurrent machines; GPU profiling is available as a trial. Pro pricing is based on machines and GPUs, but no price is listed.