René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"Querying current device is slow compared to CuPy","articleBody":"Getting the current device using `cuda.core` is quite a bit slower than CuPy:\n\n```python\nIn [1]: import cupy as cp\n\nIn [2]: %timeit cp.cuda.Device()\n69 ns ± 0.496 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)\n\nIn [3]: from cuda.core.experimental import Device\n\nIn [4]: %timeit Device()\n795 ns ± 0.273 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)\n```\n\nUltimately, my goal is to get the compute capability of the current device, and this is even slower:\n\n```python\nIn [5]: %timeit cp.cuda.Device().compute_capability\n89.1 ns ± 0.413 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)\n\nIn [6]: %timeit Device().compute_capability\n2.64 μs ± 122 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)\n```\n\nAre there tricks (e.g., caching) CuPy is employing here that `cuda.core` can use as well? Alternately, is there another way for me to use `cuda.core` or `cuda.bindings` to get this information quickly? Note that for my use case, I'm not super concerned about the _first_ call to `Device()`, but I do want _subsequent_ calls to be trivially inexpensive if the current device hasn't changed.\n\n---\n\nUsing the low-level cuda.bindings is also not quite as fast:\n\n```python\nIn [11]: def get_cc():\n    ...:     dev = runtime.cudaGetDevice()[1]\n    ...:     return driver.cuDeviceComputeCapability(dev)\n    ...:\n\nIn [12]: get_cc()\nOut[12]: (\u003cCUresult.CUDA_SUCCESS: 0>, 7, 5)\n\nIn [13]: %timeit get_cc()\n597 ns ± 0.494 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)\n```\n","author":{"url":"https://github.com/shwina","@type":"Person","name":"shwina"},"datePublished":"2025-02-07T16:17:52Z","url":"https://github.com/NVIDIA/cuda-python/issues/439"}