JAX is this rad library from Google research that acts as a Domain Specific Language (DSL) for machine learning, similar to how Halide-lang is a DSL for image processing. It’s pretty hard to get it working properly on Ubuntu since there are a lot of easy pitfalls. However, this guide will get you up and running on the latest bleeding edge of everything on Ubuntu 22.04.

First remove previous Nvidia drivers

There are two normal ways to install Nvidia drivers on Ubuntu that are familiar (1) Download the run-file from Ubuntu and manually install and (2) Use the nvidia-driver-515 package.

To uninstall the runfile version:

sudo bash --uninstall

To uninstall the Ubuntu package version:

sudo apt remove nvidia-driver-*
sudo apt autoremove

The autoremove is important to get rid of some files we need to overwrite next with symlinks.

Install cuda package

Instructions are here:

sudo dpkg -i cuda-keyring_1.0-1_all.deb
sudo apt-get update
sudo apt-get -y install cuda cuda-11-8
sudo update-alternatives --set cuda /usr/local/cuda-11.8
sudo reboot now

Check which version of CUDA is installed:

(base) catid@nuc:~$ nvidia-smi
Mon Oct 17 00:17:59 2022       
| NVIDIA-SMI 520.61.05    Driver Version: 520.61.05    CUDA Version: 11.8     |
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|   0  NVIDIA GeForce ...  On   | 00000000:01:00.0 Off |                  N/A |
|  0%   54C    P8    26W / 350W |      1MiB / 24576MiB |      0%      Default |
|                               |                      |                  N/A |
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|  No running processes found                                                 |

You can see version 11.8 of CUDA is installed. Since we ran the up-date-alternatives command, we can now use the nvcc compiler.

(base) catid@nuc:~$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Sep_21_10:33:58_PDT_2022
Cuda compilation tools, release 11.8, V11.8.89
Build cuda_11.8.r11.8/compiler.31833905_0

Test CUDA installation

git clone
cd ./cuda-samples/Samples/1_Utilities/deviceQuery
vi Makefile

Set the CUDA_PATH= line to install path of CUDA. Example:

CUDA_PATH ?= /usr/local/cuda-11.8

This path is important to note for when we install JAX later.

Build and run the tool:

(base) catid@nuc:~/cuda-samples/Samples/1_Utilities/deviceQuery$ ./deviceQuery 
./deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "NVIDIA GeForce RTX 3090"
  CUDA Driver Version / Runtime Version          11.8 / 11.8
  CUDA Capability Major/Minor version number:    8.6
  Total amount of global memory:                 24268 MBytes (25447170048 bytes)
  (082) Multiprocessors, (128) CUDA Cores/MP:    10496 CUDA Cores
  GPU Max Clock rate:                            1725 MHz (1.73 GHz)
  Memory Clock rate:                             9751 Mhz
  Memory Bus Width:                              384-bit
  L2 Cache Size:                                 6291456 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
  Maximum Layered 1D Texture Size, (num) layers  1D=(32768), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(32768, 32768), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total shared memory per multiprocessor:        102400 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  1536
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 2 copy engine(s)
  Run time limit on kernels:                     No
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  Device supports Unified Addressing (UVA):      Yes
  Device supports Managed Memory:                Yes
  Device supports Compute Preemption:            Yes
  Supports Cooperative Kernel Launch:            Yes
  Supports MultiDevice Co-op Kernel Launch:      Yes
  Device PCI Domain ID / Bus ID / location ID:   0 / 1 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 11.8, CUDA Runtime Version = 11.8, NumDevs = 1
Result = PASS

Install cuDNN

Download the .deb file from

This requires an Nvidia developer account.

I downloaded and installed this file:

sudo dpkg -i cudnn-local-repo-ubuntu2204-
sudo cp /var/cudnn-local-repo-ubuntu2204- /usr/share/keyrings/
sudo apt install libcudnn8-dev

This seems to put the cudnn library headers/files under /usr.

Easy install jax

This seems to work fine:

pip install --upgrade "jax[cuda]" -f

Manual build and install jaxlib

These steps are from

Build jaxlib:

git clone
cd jax
sudo apt install g++ python3 python3-dev
pip install numpy wheel
python build/ --enable_cuda --cuda_path /usr/local/cuda-11.8 --cudnn_path /usr

The build takes about 30 minutes on a modern processor. Finally, time to install!

pip install --upgrade pip
pip install dist/*.whl
pip install -e .  # installs jax

Test JAX

Write a simple test Python script from the Jax repo

from jax import grad
import jax.numpy as jnp

def tanh(x):  # Define a function
  y = jnp.exp(-2.0 * x)
  return (1.0 - y) / (1.0 + y)

grad_tanh = grad(tanh)  # Obtain its gradient function
print(grad_tanh(1.0))   # Evaluate it at x = 1.0
# prints 0.4199743

Run the test:

(base) catid@nuc:~$ python

I used tmux and watch nvidia-smi in a second pane to verify that the GPU was actually getting used for this script.