cudaMallocAsync / cudaMalloc different behave on different driver version

[WilliamKyle]

Code

cudaMallocAsync

#include <cuda_runtime_api.h>

#include <iostream>

using namespace std;

int main(int argc, char *argv[]) {
    void *data;
    cudaStream_t stream_ = 0;
    for (size_t i = 0; i < 32; i++) {
        cudaError_t error = cudaMallocAsync(&data, 1024 * 1024 * 1024, stream_);
        cout << "Gpu use: " << (i + 1) << "G " << cudaGetErrorString(error) << endl;
    }
    getchar();
    return 0;
}

cudaMalloc

#include <cuda_runtime_api.h>

#include <iostream>

using namespace std;

int main(int argc, char *argv[]) {
    void *data;
    cudaStream_t stream_ = 0;
    for (size_t i = 0; i < 32; i++) {
        cudaError_t error = cudaMalloc(&data, 1024 * 1024 * 1024);
        cout << "Gpu use: " << (i + 1) << "G " << cudaGetErrorString(error) << endl;
    }
    getchar();
    return 0;
}

Driver Version 536.67

No Option CUDA - Sysmem Fallback Policy

cudaMallocAsync

cudaMalloc

Driver Version 551.61

Prefer System Fallback

cudaMallocAsync

cudaMalloc

Prefer No System Fallback

cudaMallocAsync

cudaMalloc

cudaMallocAsync has one more failure when get 12G on 551.61 with Prefer System Fallback