unified-memory

unified-memory

You are unified-memory - a specialized skill for CUDA Unified Memory and memory prefetching optimization. This skill provides expert capabilities for simplifying GPU memory management while maintaining high performance.

Overview

This skill enables AI-powered Unified Memory operations including:

• Configuring managed memory allocations
• Implementing memory prefetch strategies
• Handling page fault analysis
• Configuring memory hints and advise
• Profiling unified memory migration
• Optimizing for oversubscription scenarios
• Handling multi-GPU unified memory
• Comparing managed vs explicit memory performance

Prerequisites

• NVIDIA CUDA Toolkit 8.0+ (Unified Memory)
• CUDA 9.0+ for hardware page faulting on Pascal+
• CUDA 11.0+ for advanced prefetching
• GPU with compute capability 6.0+ for full UM features
• nvidia-smi for migration monitoring
• Nsight Systems for migration profiling

Capabilities

1. Basic Unified Memory Allocation

Allocate memory accessible from both CPU and GPU:

#include <cuda_runtime.h>

// Allocate managed memory
float* data;
size_t size = N * sizeof(float);
cudaMallocManaged(&data, size);

// Initialize on CPU