GPU memory overhead for PyCUDA

gpu_memory_overhead_pycuda.py

import numpy as np
from pynvml.smi import nvidia_smi
import pycuda.gpuarray as ga
import pycuda.driver as cuda

nvsmi = nvidia_smi.getInstance()

def getGPUMemoryUsage(gpu_index=0):
    return nvsmi.DeviceQuery("memory.used")["gpu"][gpu_index]['fb_memory_usage']['used']

gpu_index = 0

print("Before: used GPU Memory: {} MB".format(getGPUMemoryUsage(gpu_index)))

cuda.init()
context = cuda.Device(gpu_index).make_context()

tensor_gpu = ga.to_gpu(np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32))

print("After: used GPU Memory: {} MB".format(getGPUMemoryUsage(gpu_index)))
context.pop()