NVIDIA Pascal CUDA Application Tuning Guide

This application note provides a guide to tuning CUDA applications for the NVIDIA Pascal compute architecture


NVIDIA Pascal CUDA Application Tuning Guide - cover page
Brand
Nvidia
Category
Other Electronics
Document type
Application Note
Language
English
Pages
12
File format
PDF
File size
1.5 MB
Published
27 October, 2023
Updated
27 October, 2023
MD5 checksum
7CF20AF8CA25482B4E8CB0C38CCA2D91

About this manual

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.

What's inside

  • Pascal Compute Architecture
  • CUDA Best Practices
  • Streaming Multiprocessor
  • Instruction Scheduling
  • Occupancy
  • New Arithmetic Primitives
  • FP16 Arithmetic Support
  • INT8 Dot Product
  • Memory Throughput
  • Unified L1/Texture Cache

Specifications

Compute Capability6.x
GP104 FP32 Cores per SM128
GP104 FP64 Cores per SM4
GP100 SM Countup to 60
GP100 FP32 Cores per SM64
GP100 FP64 Cores per SM32
GP100 Shared Memory per SM64KB
GP104 Shared Memory per SM96KB
GP100 Peak Bandwidth732 GB/s
GP100 Memory Clock715 MHz
GP100 FP16 Throughput2x FP32
GP104 FP16 Throughput1/64th FP32
Max Registers per Thread255
Max Thread Blocks per SM32
Max Concurrent Warps per SM64
Max Shared Memory per Block48KB

Frequently asked questions

What are the Pascal compute capabilities?

Pascal refers to devices of compute capability 6.x, with GP100 at 6.0 and GP104 at 6.1.

How does GP100 FP16 throughput compare to FP32?

GP100 provides FP16 throughput up to 2x that of FP32 arithmetic.

What is the peak bandwidth of the GP100 equipped Tesla P100?

The GP100 equipped Tesla P100 has a peak bandwidth of 732 GB/s with a 715 MHz memory clock.

How can developers enable caching of global loads in the unified L1/Texture cache on GP104?

Developers can opt-in by passing the -Xptxas -dlcm=ca flag to nvcc at compile time.

Download PDF (1.5 MB)

More Nvidia Manuals


View all Nvidia manuals →