This application note provides a guide to tuning CUDA applications for the NVIDIA Pascal compute architecture
This document serves as a tuning guide for CUDA applications running on the NVIDIA Pascal architecture, which includes the GP100 and GP104 variants. It outlines how developers can optimize their applications to leverage specific architectural features of Pascal for improved performance without requiring code changes, building upon best practices established for previous architectures like Kepler and Maxwell.
| Compute Capability | 6.x |
|---|---|
| GP104 FP32 Cores per SM | 128 |
| GP104 FP64 Cores per SM | 4 |
| GP100 SM Count | up to 60 |
| GP100 FP32 Cores per SM | 64 |
| GP100 FP64 Cores per SM | 32 |
| GP100 Shared Memory per SM | 64KB |
| GP104 Shared Memory per SM | 96KB |
| GP100 Peak Bandwidth | 732 GB/s |
| GP100 Memory Clock | 715 MHz |
| GP100 FP16 Throughput | 2x FP32 |
| GP104 FP16 Throughput | 1/64th FP32 |
| Max Registers per Thread | 255 |
| Max Thread Blocks per SM | 32 |
| Max Concurrent Warps per SM | 64 |
| Max Shared Memory per Block | 48KB |
Pascal refers to devices of compute capability 6.x, with GP100 at 6.0 and GP104 at 6.1.
GP100 provides FP16 throughput up to 2x that of FP32 arithmetic.
The GP100 equipped Tesla P100 has a peak bandwidth of 732 GB/s with a 715 MHz memory clock.
Developers can opt-in by passing the -Xptxas -dlcm=ca flag to nvcc at compile time.