Introduction

TensorFlow 1.15 is old enough that installation problems are usually compatibility problems, not Conda problems. The key is to match a Python version and GPU runtime that TensorFlow 1.15 actually supports, then keep everything isolated inside one Conda environment.

For the classic GPU setup, the safest target is usually Python 3.7, cudatoolkit=10.0, and cudnn=7.6, with the TensorFlow 1.15 GPU package installed into that environment. Newer NVIDIA drivers are fine as long as they are compatible with the CUDA 10 runtime.

Build a Clean Conda Environment First

Start from a fresh environment so older TensorFlow, CUDA, or NumPy packages do not leak in:

conda create -n tf115 python=3.7 -y
conda activate tf115

Using a dedicated environment matters because TensorFlow 1.15 often conflicts with modern package stacks. Trying to drop it into an existing machine-learning environment usually creates hard-to-debug library mismatches.

Install CUDA and cuDNN Through Conda

For a Conda-based install, you normally do not need the full standalone CUDA toolkit installed system-wide. You do still need a working NVIDIA driver on the machine, but Conda can provide the CUDA runtime libraries used by TensorFlow.

Install the historical versions that match the TensorFlow 1.15 GPU line:

conda install -y cudatoolkit=10.0 cudnn=7.6

This is the most important step. If you install a newer CUDA runtime such as 11.x into the environment and expect TensorFlow 1.15 to use it, import failures are very common.

Install TensorFlow 1.15

Now install the TensorFlow package itself:

pip install "tensorflow-gpu==1.15.5"

If you specifically need plain 1.15.0 rather than the last 1.15.x patch release, change the version accordingly. In most cases, 1.15.5 is the better choice because it stays on the same compatibility line while picking up the later patch fixes.

At this point, the environment usually contains everything TensorFlow needs except the host NVIDIA driver.

Verify the Installation

Test that TensorFlow imports correctly and can see the GPU:

1python - <<'PY'
2import tensorflow as tf
3print(tf.__version__)
4print(tf.test.is_gpu_available())
5print(tf.test.gpu_device_name())
6PY

A successful result should show version 1.15.x and report a GPU as available. If the import works but GPU detection is false, the usual causes are:

• unsupported CUDA or cuDNN version in the Conda environment
• NVIDIA driver missing or too old
• another library path overriding the Conda runtime

A Minimal End-to-End Recipe

A compact working sequence looks like this:

1conda create -n tf115 python=3.7 -y
2conda activate tf115
3conda install -y cudatoolkit=10.0 cudnn=7.6
4pip install "tensorflow-gpu==1.15.5"
5python -c "import tensorflow as tf; print(tf.__version__); print(tf.test.is_gpu_available())"

That is the setup many legacy training pipelines still expect.

Conda Versus System CUDA

One confusing point is whether you need /usr/local/cuda or a full CUDA toolkit installer from NVIDIA. With a Conda-driven TensorFlow 1.15 environment, the answer is usually no. Conda supplies the runtime libraries, and the host machine only needs the NVIDIA driver.

Problems appear when people mix approaches. For example, a system-wide CUDA 11 install can end up on the library path while TensorFlow 1.15 is expecting CUDA 10.0 from the Conda environment. When that happens, import errors and missing-symbol messages become likely.

If you must use system CUDA for some reason, be very deliberate about library paths.

Common Pitfalls

The biggest pitfall is using an unsupported Python version. TensorFlow 1.15 is from an older Python ecosystem, so a modern default like Python 3.11 is not going to work.

Another common issue is installing the wrong CUDA line. TensorFlow 1.15 is tied to the older CUDA 10.0 and cuDNN 7.x generation, not the modern 11.x or 12.x stack.

People also mix Conda-provided GPU libraries with system-wide CUDA libraries and end up loading the wrong .so or .dll files at runtime.

Finally, do not forget the host driver. Conda can provide cudatoolkit and cudnn, but it cannot replace the NVIDIA driver that actually talks to the hardware.

Summary

• TensorFlow 1.15 works best in a dedicated Conda environment with older compatible dependencies.
• A common working combination is Python 3.7, cudatoolkit=10.0, and cudnn=7.6.
• Use tensorflow-gpu==1.15.5 unless you have a strict reason to pin exactly 1.15.0.
• Conda usually removes the need for a full system CUDA toolkit install, but you still need a valid NVIDIA driver.
• Most failures come from mismatched Python, CUDA, cuDNN, or mixed system-versus-Conda library paths.