From Beocat
Revision as of 14:25, 16 October 2017 by Kylehutson (talk | contribs) (Changed info to deal with updates from Paladins to Dwarves)

Jump to: navigation, search

CUDA Overview

CUDA is a feature set for programming nVidia GPUs. We have 7 CUDA-enabled nodes. dwarf22, dwarf23, dwarf24, dwarf25, and dwarf35 each have two nVidia 1080 Ti graphics cards. dwarf38 and dwarf39 each have a single nVidia 980 Ti graphic card. The former set of nodes is only available to 'killable" jobs for those outside the research group that purchased them. The latter are available for anybody, however, you should send an email to with a request to be added to the GPU priority group.

Note that both of these graphic cards are consumer-grade rather than the typical GPUs used in most high-performance computing centers. For single-precision computations, these cards are comparable to the high-end cards (at a fraction of the price), however double-precision computations are much slower.

Training videos

CUDA Programming Model Overview:

CUDA Programming Basics Part I (Host functions):

CUDA Programming Basics Part II (Device functions):

Compiling CUDA Applications

nvcc is the compiler for CUDA applications. When compiling your applications manually you will need to keep 3 things in mind:

  • The CUDA development headers are located here: /opt/cuda/sdk/common/inc
  • The CUDA architecture is: sm_30
  • The CUDA SDK is currently not available on the headnode. (compile on the nodes with CUDA, either in your jobscript or via qrsh -l cuda=TRUE)
  • Do not run your cuda applications on the headnode. I cannot guarantee it will run, and it will give you terrible results if it does run.

Putting it all together you can compile CUDA applications as follows:

nvcc -I  /opt/cuda/sdk/common/inc -arch sm_30 <source>.cu -o <output>


Create your Application

Copy the following Application into Beocat as

//  Kernel definition, see also section 4.2.3 of Nvidia Cuda Programming Guide
__global__  void vecAdd(float* A, float* B, float* C)
            // threadIdx.x is a built-in variable  provided by CUDA at runtime
            int i = threadIdx.x;
       C[i] = A[i] + B[i];

#include  <stdio.h>
#define  SIZE 10
int  main()
   int N=SIZE;
   float A[SIZE], B[SIZE], C[SIZE];
   float *devPtrA;
   float *devPtrB;
   float *devPtrC;
   int memsize= SIZE * sizeof(float);

   cudaMalloc((void**)&devPtrA, memsize);
   cudaMalloc((void**)&devPtrB, memsize);
   cudaMalloc((void**)&devPtrC, memsize);
   cudaMemcpy(devPtrA, A, memsize,  cudaMemcpyHostToDevice);
   cudaMemcpy(devPtrB, B, memsize,  cudaMemcpyHostToDevice);
   // __global__ functions are called:  Func<<< Dg, Db, Ns  >>>(parameter);
   vecAdd<<<1, N>>>(devPtrA,  devPtrB, devPtrC);
   cudaMemcpy(C, devPtrC, memsize,  cudaMemcpyDeviceToHost);

   for (int i=0; i<SIZE; i++)



Gain Access to a CUDA-capable Node

qrsh -l cuda=TRUE

Compile Your Application

nvcc -I /opt/cuda/sdk/common/inc -arch sm_30 -o vecadd

This will create a program with the name 'vecadd' (specified by the '-o' flag).

Run Your Application

Run the program as you usually would, namely


Assuming you don't want to run the program interactively because this is a large job, you can submit a job via qsub, just be sure to add the '-l cuda=true' directive.