Simple python script to obtain CUDA device information

cuda_check.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

"""
Outputs some information on CUDA-enabled devices on your computer,
including current memory usage.

It's a port of https://gist.github.com/f0k/0d6431e3faa60bffc788f8b4daa029b1
from C to Python with ctypes, so it can run without compiling anything. Note
that this is a direct translation with no attempt to make the code Pythonic.
It's meant as a general demonstration on how to obtain CUDA device information
from Python without resorting to nvidia-smi or a compiled Python extension.

Author: Jan Schlüter
License: MIT (https://gist.github.com/f0k/63a664160d016a491b2cbea15913d549#gistcomment-3870498)
"""

import sys
import ctypes


# Some constants taken from cuda.h
CUDA_SUCCESS = 0
CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR = 39
CU_DEVICE_ATTRIBUTE_CLOCK_RATE = 13
CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE = 36


def ConvertSMVer2Cores(major, minor):
    # Returns the number of CUDA cores per multiprocessor for a given
    # Compute Capability version. There is no way to retrieve that via
    # the API, so it needs to be hard-coded.
    # See _ConvertSMVer2Cores in helper_cuda.h in NVIDIA's CUDA Samples.
    return {(1, 0): 8,    # Tesla
            (1, 1): 8,
            (1, 2): 8,
            (1, 3): 8,
            (2, 0): 32,   # Fermi
            (2, 1): 48,
            (3, 0): 192,  # Kepler
            (3, 2): 192,
            (3, 5): 192,
            (3, 7): 192,
            (5, 0): 128,  # Maxwell
            (5, 2): 128,
            (5, 3): 128,
            (6, 0): 64,   # Pascal
            (6, 1): 128,
            (6, 2): 128,
            (7, 0): 64,   # Volta
            (7, 2): 64,
            (7, 5): 64,   # Turing
            (8, 0): 64,   # Ampere
            (8, 6): 128,
            (8, 7): 128,
            (8, 9): 128,  # Ada
            (9, 0): 128,  # Hopper
            }.get((major, minor), 0)


def main():
    libnames = ('libcuda.so', 'libcuda.dylib', 'nvcuda.dll', 'cuda.dll')
    for libname in libnames:
        try:
            cuda = ctypes.CDLL(libname)
        except OSError:
            continue
        else:
            break
    else:
        raise OSError("could not load any of: " + ' '.join(libnames))

    nGpus = ctypes.c_int()
    name = b' ' * 100
    cc_major = ctypes.c_int()
    cc_minor = ctypes.c_int()
    cores = ctypes.c_int()
    threads_per_core = ctypes.c_int()
    clockrate = ctypes.c_int()
    freeMem = ctypes.c_size_t()
    totalMem = ctypes.c_size_t()

    result = ctypes.c_int()
    device = ctypes.c_int()
    context = ctypes.c_void_p()
    error_str = ctypes.c_char_p()

    result = cuda.cuInit(0)
    if result != CUDA_SUCCESS:
        cuda.cuGetErrorString(result, ctypes.byref(error_str))
        print("cuInit failed with error code %d: %s" % (result, error_str.value.decode()))
        return 1
    result = cuda.cuDeviceGetCount(ctypes.byref(nGpus))
    if result != CUDA_SUCCESS:
        cuda.cuGetErrorString(result, ctypes.byref(error_str))
        print("cuDeviceGetCount failed with error code %d: %s" % (result, error_str.value.decode()))
        return 1
    print("Found %d device(s)." % nGpus.value)
    for i in range(nGpus.value):
        result = cuda.cuDeviceGet(ctypes.byref(device), i)
        if result != CUDA_SUCCESS:
            cuda.cuGetErrorString(result, ctypes.byref(error_str))
            print("cuDeviceGet failed with error code %d: %s" % (result, error_str.value.decode()))
            return 1
        print("Device: %d" % i)
        if cuda.cuDeviceGetName(ctypes.c_char_p(name), len(name), device) == CUDA_SUCCESS:
            print("  Name: %s" % (name.split(b'\0', 1)[0].decode()))
        if cuda.cuDeviceComputeCapability(ctypes.byref(cc_major), ctypes.byref(cc_minor), device) == CUDA_SUCCESS:
            print("  Compute Capability: %d.%d" % (cc_major.value, cc_minor.value))
        if cuda.cuDeviceGetAttribute(ctypes.byref(cores), CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device) == CUDA_SUCCESS:
            print("  Multiprocessors: %d" % cores.value)
            print("  CUDA Cores: %s" % (cores.value * ConvertSMVer2Cores(cc_major.value, cc_minor.value) or "unknown"))
            if cuda.cuDeviceGetAttribute(ctypes.byref(threads_per_core), CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR, device) == CUDA_SUCCESS:
                print("  Concurrent threads: %d" % (cores.value * threads_per_core.value))
        if cuda.cuDeviceGetAttribute(ctypes.byref(clockrate), CU_DEVICE_ATTRIBUTE_CLOCK_RATE, device) == CUDA_SUCCESS:
            print("  GPU clock: %g MHz" % (clockrate.value / 1000.))
        if cuda.cuDeviceGetAttribute(ctypes.byref(clockrate), CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE, device) == CUDA_SUCCESS:
            print("  Memory clock: %g MHz" % (clockrate.value / 1000.))
        try:
            result = cuda.cuCtxCreate_v2(ctypes.byref(context), 0, device)
        except AttributeError:
            result = cuda.cuCtxCreate(ctypes.byref(context), 0, device)
        if result != CUDA_SUCCESS:
            cuda.cuGetErrorString(result, ctypes.byref(error_str))
            print("cuCtxCreate failed with error code %d: %s" % (result, error_str.value.decode()))
        else:
            try:
                result = cuda.cuMemGetInfo_v2(ctypes.byref(freeMem), ctypes.byref(totalMem))
            except AttributeError:
                result = cuda.cuMemGetInfo(ctypes.byref(freeMem), ctypes.byref(totalMem))
            if result == CUDA_SUCCESS:
                print("  Total Memory: %ld MiB" % (totalMem.value / 1024**2))
                print("  Free Memory: %ld MiB" % (freeMem.value / 1024**2))
            else:
                cuda.cuGetErrorString(result, ctypes.byref(error_str))
                print("cuMemGetInfo failed with error code %d: %s" % (result, error_str.value.decode()))
            cuda.cuCtxDetach(context)
    return 0


if __name__=="__main__":
    sys.exit(main())

https://gist.github.com/f0k/63a664160d016a491b2cbea15913d549#file-cuda_check-py-L60
I believe it should look for nvcuda.dll here on windows.
As i know cuda.dll is deprecated since 1.1 version of cuda and then replaced by nvcuda.dll which distributed with Nvidia Driver since 169.x.x version. I test it localy and it works. I am not an expert though.

This is Cuda programming code and Compare GPU vs CPU

https://debuggingsolution.blogspot.com/2021/09/vector-addition-cuda-parallel.html

Thanks @f0k for the excellent snippet! Here is an importable version which can be run inside other scripts as get_cuda_device_specs(). It returns a list of specification dicts per CUDA device

[
  {
    "name": "NVIDIA GeForce RTX 3080 Laptop GPU",
    "compute_capability": [
      8,
      6
    ],
    "architecture": "ampere",
    "cores": 48,
    "cuda_cores": 3072,
    "concurrent_threads": 73728,
    "gpu_clock_mhz": 1245.0,
    "mem_clock_mhz": 6001.0,
    "total_mem_mb": 16125.3125,
    "free_mem_mb": 15733.25
  }
]

I also made some minor cosmetic updates

• Refactor str.format() to f-strings for readability
• Refactor camel case to snake case (for PEP linting)
• Move semantic versioning map to constant named dict
• Add another mapping to the architecture key name
• Switch sys.exit codes to RuntimeError and warnings.warn where appropriate

import ctypes
import json
from typing import Any, Dict, List
from warnings import warn

# TODO define decorator to share the RuntimeError/CUDA_SUCCESS logic among different library functions

# One of the following libraries must be available to load
libnames = ('libcuda.so', 'libcuda.dylib', 'cuda.dll')
for libname in libnames:
    try:
        cuda = ctypes.CDLL(libname)
    except OSError:
        continue
    else:
        break
else:
    raise ImportError(f'Could not load any of: {", ".join(libnames)}')

# Constants from cuda.h
CUDA_SUCCESS = 0
CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR = 39
CU_DEVICE_ATTRIBUTE_CLOCK_RATE = 13
CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE = 36

# Conversions from semantic version numbers
# Borrowed from original gist and updated from the "GPUs supported" section of this Wikipedia article
# https://en.wikipedia.org/wiki/CUDA
SEMVER_TO_CORES = {
    (1, 0): 8,    # Tesla
    (1, 1): 8,
    (1, 2): 8,
    (1, 3): 8,
    (2, 0): 32,   # Fermi
    (2, 1): 48,
    (3, 0): 192,  # Kepler
    (3, 2): 192,
    (3, 5): 192,
    (3, 7): 192,
    (5, 0): 128,  # Maxwell
    (5, 2): 128,
    (5, 3): 128,
    (6, 0): 64,   # Pascal
    (6, 1): 128,
    (6, 2): 128,
    (7, 0): 64,   # Volta
    (7, 2): 64,
    (7, 5): 64,   # Turing
    (8, 0): 64,   # Ampere
    (8, 6): 64,
}
SEMVER_TO_ARCH = {
    (1, 0): 'tesla',
    (1, 1): 'tesla',
    (1, 2): 'tesla',
    (1, 3): 'tesla',

    (2, 0): 'fermi',
    (2, 1): 'fermi',

    (3, 0): 'kepler',
    (3, 2): 'kepler',
    (3, 5): 'kepler',
    (3, 7): 'kepler',

    (5, 0): 'maxwell',
    (5, 2): 'maxwell',
    (5, 3): 'maxwell',

    (6, 0): 'pascal',
    (6, 1): 'pascal',
    (6, 2): 'pascal',

    (7, 0): 'volta',
    (7, 2): 'volta',

    (7, 5): 'turing',

    (8, 0): 'ampere',
    (8, 6): 'ampere',
}


def get_cuda_device_specs() -> List[Dict[str, Any]]:
    """Generate spec for each GPU device with format

    {
        'name': str,
        'compute_capability': (major: int, minor: int),
        'cores': int,
        'cuda_cores': int,
        'concurrent_threads': int,
        'gpu_clock_mhz': float,
        'mem_clock_mhz': float,
        'total_mem_mb': float,
        'free_mem_mb': float
    }
    """

    # Type-binding definitions for ctypes
    num_gpus = ctypes.c_int()
    name = b' ' * 100
    cc_major = ctypes.c_int()
    cc_minor = ctypes.c_int()
    cores = ctypes.c_int()
    threads_per_core = ctypes.c_int()
    clockrate = ctypes.c_int()
    free_mem = ctypes.c_size_t()
    total_mem = ctypes.c_size_t()
    result = ctypes.c_int()
    device = ctypes.c_int()
    context = ctypes.c_void_p()
    error_str = ctypes.c_char_p()

    # Check expected initialization
    result = cuda.cuInit(0)
    if result != CUDA_SUCCESS:
        cuda.cuGetErrorString(result, ctypes.byref(error_str))
        raise RuntimeError(f'cuInit failed with error code {result}: {error_str.value.decode()}')
    result = cuda.cuDeviceGetCount(ctypes.byref(num_gpus))
    if result != CUDA_SUCCESS:
        cuda.cuGetErrorString(result, ctypes.byref(error_str))
        raise RuntimeError(f'cuDeviceGetCount failed with error code {result}: {error_str.value.decode()}')

    # Iterate through available devices
    device_specs = []
    for i in range(num_gpus.value):
        spec = {}
        result = cuda.cuDeviceGet(ctypes.byref(device), i)
        if result != CUDA_SUCCESS:
            cuda.cuGetErrorString(result, ctypes.byref(error_str))
            raise RuntimeError(f'cuDeviceGet failed with error code {result}: {error_str.value.decode()}')

        # Parse specs for each device
        if cuda.cuDeviceGetName(ctypes.c_char_p(name), len(name), device) == CUDA_SUCCESS:
            spec.update(name=name.split(b'\0', 1)[0].decode())
        if cuda.cuDeviceComputeCapability(ctypes.byref(cc_major), ctypes.byref(cc_minor), device) == CUDA_SUCCESS:
            spec.update(compute_capability=(cc_major.value, cc_minor.value))
            spec.update(architecture=SEMVER_TO_ARCH.get((cc_major.value, cc_minor.value), 'unknown'))
        if cuda.cuDeviceGetAttribute(ctypes.byref(cores), CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device) == CUDA_SUCCESS:
            spec.update(
                cores=cores.value,
                cuda_cores=cores.value * SEMVER_TO_CORES.get((cc_major.value, cc_minor.value), 'unknown'))
            if cuda.cuDeviceGetAttribute(ctypes.byref(threads_per_core), CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR, device) == CUDA_SUCCESS:
                spec.update(concurrent_threads=cores.value * threads_per_core.value)
        if cuda.cuDeviceGetAttribute(ctypes.byref(clockrate), CU_DEVICE_ATTRIBUTE_CLOCK_RATE, device) == CUDA_SUCCESS:
            spec.update(gpu_clock_mhz=clockrate.value / 1000.)
        if cuda.cuDeviceGetAttribute(ctypes.byref(clockrate), CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE, device) == CUDA_SUCCESS:
            spec.update(mem_clock_mhz=clockrate.value / 1000.)

        # Attempt to determine available vs. free memory
        try:
            result = cuda.cuCtxCreate_v2(ctypes.byref(context), 0, device)
        except AttributeError:
            result = cuda.cuCtxCreate(ctypes.byref(context), 0, device)
        if result != CUDA_SUCCESS:
            cuda.cuGetErrorString(result, ctypes.byref(error_str))
            warn(f'cuCtxCreate failed with error code {result}: {error_str.value.decode()}')
        else:
            try:
                result = cuda.cuMemGetInfo_v2(ctypes.byref(free_mem), ctypes.byref(total_mem))
            except AttributeError:
                result = cuda.cuMemGetInfo(ctypes.byref(free_mem), ctypes.byref(total_mem))
            if result == CUDA_SUCCESS:
                spec.update(
                    total_mem_mb=total_mem.value / 1024**2,
                    free_mem_mb=free_mem.value / 1024**2)
            else:
                cuda.cuGetErrorString(result, ctypes.byref(error_str))
                warn(f'cuMemGetInfo failed with error code {result}: {error_str.value.decode()}')
            cuda.cuCtxDetach(context)
        device_specs.append(spec)
    return device_specs


if __name__ == '__main__':

    print(json.dumps(get_cuda_device_specs(), indent=2))

Thank you for this script! It helped me debug an issue with getting CUDA working in Windows 10 with Ubuntu WSL.

See this: TimDettmers/bitsandbytes#337 - Thanks again @f0k !!!

Thanks! Further refactoring with decorators.

import ctypes
import json
from functools import wraps
from typing import Any, Dict, List
from warnings import warn

# Constants from cuda.h
CUDA_SUCCESS = 0
CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR = 39
CU_DEVICE_ATTRIBUTE_CLOCK_RATE = 13
CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE = 36

# Conversions from semantic version numbers
# Borrowed from original gist and updated from the "GPUs supported" section of this Wikipedia article
# https://en.wikipedia.org/wiki/CUDA
SEMVER_TO_CORES = {
    (1, 0): 8,  # Tesla
    (1, 1): 8,
    (1, 2): 8,
    (1, 3): 8,
    (2, 0): 32,  # Fermi
    (2, 1): 48,
    (3, 0): 192,  # Kepler
    (3, 2): 192,
    (3, 5): 192,
    (3, 7): 192,
    (5, 0): 128,  # Maxwell
    (5, 2): 128,
    (5, 3): 128,
    (6, 0): 64,  # Pascal
    (6, 1): 128,
    (6, 2): 128,
    (7, 0): 64,  # Volta
    (7, 2): 64,
    (7, 5): 64,  # Turing
    (8, 0): 64,  # Ampere
    (8, 6): 64,
}
SEMVER_TO_ARCH = {
    (1, 0): "tesla",
    (1, 1): "tesla",
    (1, 2): "tesla",
    (1, 3): "tesla",
    (2, 0): "fermi",
    (2, 1): "fermi",
    (3, 0): "kepler",
    (3, 2): "kepler",
    (3, 5): "kepler",
    (3, 7): "kepler",
    (5, 0): "maxwell",
    (5, 2): "maxwell",
    (5, 3): "maxwell",
    (6, 0): "pascal",
    (6, 1): "pascal",
    (6, 2): "pascal",
    (7, 0): "volta",
    (7, 2): "volta",
    (7, 5): "turing",
    (8, 0): "ampere",
    (8, 6): "ampere",
}


# Decorator for CUDA API calls
def cuda_api_call(func):
    """
    Decorator to wrap CUDA API calls and check their results.
    Raises RuntimeError if the CUDA call does not return CUDA_SUCCESS.
    """

    @wraps(func)
    def wrapper(*args, **kwargs):
        result = func(*args, **kwargs)
        if result != CUDA_SUCCESS:
            error_str = ctypes.c_char_p()
            cuda.cuGetErrorString(result, ctypes.byref(error_str))
            raise RuntimeError(
                f"{func.__name__} failed with error code {result}: {error_str.value.decode()}"
            )
        return result

    return wrapper


def cuda_api_call_warn(func):
    """
    Decorator to wrap CUDA API calls and check their results.
    Prints a warning message if the CUDA call does not return CUDA_SUCCESS.
    """

    @wraps(func)
    def wrapper(*args, **kwargs):
        result = func(*args, **kwargs)
        if result != CUDA_SUCCESS:
            error_str = ctypes.c_char_p()
            cuda.cuGetErrorString(result, ctypes.byref(error_str))
            warn(
                f"Warning: {func.__name__} failed with error code {result}: {error_str.value.decode()}"
            )
        return result

    return wrapper


# Attempt to load the CUDA library
libnames = ("libcuda.so", "libcuda.dylib", "cuda.dll")
for libname in libnames:
    try:
        cuda = ctypes.CDLL(libname)
    except OSError:
        continue
    else:
        break
else:
    raise ImportError(f'Could not load any of: {", ".join(libnames)}')


# CUDA API calls wrapped with the decorator
@cuda_api_call
def cuInit(flags):
    return cuda.cuInit(flags)


@cuda_api_call
def cuDeviceGetCount(count):
    return cuda.cuDeviceGetCount(count)


@cuda_api_call
def cuDeviceGet(device, ordinal):
    return cuda.cuDeviceGet(device, ordinal)


@cuda_api_call
def cuDeviceGetName(name, len, dev):
    return cuda.cuDeviceGetName(name, len, dev)


@cuda_api_call
def cuDeviceComputeCapability(major, minor, dev):
    return cuda.cuDeviceComputeCapability(major, minor, dev)


@cuda_api_call
def cuDeviceGetAttribute(pi, attrib, dev):
    return cuda.cuDeviceGetAttribute(pi, attrib, dev)


@cuda_api_call_warn
def cuCtxCreate(pctx, flags, dev):
    try:
        result = cuda.cuCtxCreate_v2(pctx, flags, dev)
    except AttributeError:
        result = cuda.cuCtxCreate(pctx, flags, dev)
    return result


@cuda_api_call_warn
def cuMemGetInfo(free, total):
    try:
        result = cuda.cuMemGetInfo_v2(free, total)
    except AttributeError:
        result = cuda.cuMemGetInfo(free, total)
    return result


@cuda_api_call
def cuCtxDetach(ctx):
    return cuda.cuCtxDetach(ctx)


# Main function to get CUDA device specs
def get_cuda_device_specs() -> List[Dict[str, Any]]:
    """Generate spec for each GPU device with format
    {
        'name': str,
        'compute_capability': (major: int, minor: int),
        'cores': int,
        'cuda_cores': int,
        'concurrent_threads': int,
        'gpu_clock_mhz': float,
        'mem_clock_mhz': float,
        'total_mem_mb': float,
        'free_mem_mb': float,
        'architecture': str,
        'cuda_cores': int
    }
    """
    # Initialize CUDA
    cuInit(0)

    num_gpus = ctypes.c_int()
    cuDeviceGetCount(ctypes.byref(num_gpus))

    device_specs = []
    for i in range(num_gpus.value):
        spec = {}
        device = ctypes.c_int()
        cuDeviceGet(ctypes.byref(device), i)

        name = b" " * 100
        cuDeviceGetName(ctypes.c_char_p(name), len(name), device)
        spec["name"] = name.split(b"\0", 1)[0].decode()

        cc_major = ctypes.c_int()
        cc_minor = ctypes.c_int()
        cuDeviceComputeCapability(
            ctypes.byref(cc_major), ctypes.byref(cc_minor), device
        )
        compute_capability = (cc_major.value, cc_minor.value)
        spec["compute_capability"] = compute_capability

        cores = ctypes.c_int()
        cuDeviceGetAttribute(
            ctypes.byref(cores), CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device
        )
        spec["cores"] = cores.value

        threads_per_core = ctypes.c_int()
        cuDeviceGetAttribute(
            ctypes.byref(threads_per_core),
            CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR,
            device,
        )
        spec["concurrent_threads"] = cores.value * threads_per_core.value

        clockrate = ctypes.c_int()
        cuDeviceGetAttribute(
            ctypes.byref(clockrate), CU_DEVICE_ATTRIBUTE_CLOCK_RATE, device
        )
        spec["gpu_clock_mhz"] = clockrate.value / 1000.0

        cuDeviceGetAttribute(
            ctypes.byref(clockrate), CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE, device
        )
        spec["mem_clock_mhz"] = clockrate.value / 1000.0

        context = ctypes.c_void_p()
        if cuCtxCreate(ctypes.byref(context), 0, device) == CUDA_SUCCESS:
            free_mem = ctypes.c_size_t()
            total_mem = ctypes.c_size_t()

            cuMemGetInfo(ctypes.byref(free_mem), ctypes.byref(total_mem))

            spec["total_mem_mb"] = total_mem.value / 1024**2
            spec["free_mem_mb"] = free_mem.value / 1024**2

            spec["architecture"] = SEMVER_TO_ARCH.get(compute_capability, "unknown")
            spec["cuda_cores"] = cores.value * SEMVER_TO_CORES.get(
                compute_capability, "unknown"
            )

            cuCtxDetach(context)

        device_specs.append(spec)
    return device_specs


if __name__ == "__main__":
    print(json.dumps(get_cuda_device_specs(), indent=2))