Contents

White paper covering the most common issues related to NVIDIA GPUs.

• 1. Pascal Tuning Guide
  • 1.1. NVIDIA Pascal Compute Architecture
  • 1.2. CUDA Best Practices
  • 1.3. Application Compatibility
  • 1.4. Pascal Tuning
    • 1.4.1. Streaming Multiprocessor
      • 1.4.1.1. Instruction Scheduling
      • 1.4.1.2. Occupancy
    • 1.4.2. New Arithmetic Primitives
      • 1.4.2.1. FP16 Arithmetic Support
      • 1.4.2.2. INT8 Dot Product
    • 1.4.3. Memory Throughput
      • 1.4.3.1. High Bandwidth Memory 2 DRAM
      • 1.4.3.2. Unified L1/Texture Cache
    • 1.4.4. Atomic Memory Operations
    • 1.4.5. Shared Memory
      • 1.4.5.1. Shared Memory Capacity
      • 1.4.5.2. Shared Memory Bandwidth
    • 1.4.6. Inter-GPU Communication
      • 1.4.6.1. NVLink Interconnect
      • 1.4.6.2. GPUDirect RDMA Bandwidth
    • 1.4.7. Compute Preemption
    • 1.4.8. Unified Memory Improvements
• 2. Revision History
• 3. Notices
  • 3.1. Notice
  • 3.2. OpenCL
  • 3.3. Trademarks