compute capability

**Compute capability** is the **GPU architecture version identifier that defines supported instructions, memory features, and performance behaviors** - it determines what low-level optimizations and precision modes are available to compiled CUDA kernels. **What Is Compute capability?** - **Definition**: SM version number used by CUDA toolchains to target architecture-specific features. - **Feature Envelope**: Controls availability of tensor instructions, cache behavior, and precision formats. - **Compilation Impact**: Binary generation and PTX compatibility depend on selected architecture targets. - **Runtime Effect**: Different capabilities can change kernel performance characteristics significantly. **Why Compute capability Matters** - **Correctness**: Using unsupported instructions for a target architecture causes build or runtime failures. - **Performance**: Architecture-tuned kernels can unlock major speedups over generic builds. - **Portability Planning**: Multi-architecture deployments need deliberate build matrices and compatibility policy. - **Feature Adoption**: New precision modes and acceleration paths arrive with newer compute capabilities. - **Lifecycle Management**: Capability awareness guides hardware upgrade and software roadmap decisions. **How It Is Used in Practice** - **Build Targeting**: Compile with explicit architecture flags matching deployed GPU fleets. - **Fallback Strategy**: Provide compatible kernels or binaries for older capabilities where required. - **Regression Testing**: Validate performance and numerics across each supported compute capability tier. Compute capability is **the hardware contract for CUDA software behavior** - architecture-aware builds are necessary to achieve both compatibility and peak GPU performance.