In addition to Anaconda Notebook’s default environments, Anaconda provides access to remote GPU kernels powered by NVIDIA’s A10 GPUs. These kernels are ideal for complex, compute-intensive workloads that would otherwise be too time consuming to run in a default environment. You can choose to run the entire notebook on GPU or apply GPU acceleration to specific cells.
Anaconda’s GPU kernel runs a GPU-enabled environment called “anaconda-01 - nvidia-a10”. This environment contains the libraries, drivers, and configurations necessary for communicating with NVIDIA GPUs, as well as the ipykernel package to enable interactive computing in Python. See below for a list of all packages included in the default environment.
When you assign a GPU kernel to a notebook (or notebook cell), you’re using the default GPU-enabled environment. For more information about kernels and environment, see Environments.
At this time, Anaconda is offering free credits to select paid subscribers only. Free tier subscribers who would like to use GPU kernels can upgrade to gain access.
Once you assign a GPU kernel to a notebook or cell, you can run your code as you normally would. You should notice a considerable increase in processing speed, especially for more complex tasks.
Anaconda’s default GPU kernel, “anaconda-01 - nvidia-a10”, is tied to a read-only GPU-enabled environment that includes the latest Anaconda, AI, and NVIDIA packages.
Because cells assigned a GPU kernel run in a separate environment from the rest of the notebook, variables must be transferred between the cell and the notebook to ensure they are accessible in both environments.
Follow these steps to transfer variables:
Click Variable transfer beside a cell with a GPU kernel assigned.
Enter the inputs and/or outputs to be transferred between the notebook and cell environments.
Click Set variables.
The inputs and/or outputs can now be used in the GPU kernel cell and notebook.
When you assign a GPU kernel to a notebook, you are asked to reserve a defined number of minutes to use toward the notebook’s GPU execution time. Reserved time is calculated in credits, with one credit equaling 100 real-world seconds.
Once the notebook launches, the reserved time starts to count down. The reserved time continues to count down both when running code and when the notebook is idle.
Track how much reserved time is left by hovering over the top-right countdown bar:
When your reserved time drops below five minutes (300 seconds), the countdown bar turns red. Once the reserved time hits zero, the GPU kernel session automatically terminates.
Anaconda recommends conserving your GPU usage by either assigning a GPU kernel to individual cells that require higher compute power rather than the entire notebook, or waiting to assign the GPU kernel until you’re ready to run your code, and then stopping the GPU kernel session as soon as the code has executed.
There are multiple ways you can stop a GPU kernel session:
View your GPU usage stats and remaining credits by clicking the top-right actions dropdown on the Jupyter Kernels home page, then selecting Usage:
The Overview page shows your running GPU kernels, remaining credits, and historical credit consumption.
The History page shows your usage history, including GPU session kernel names, types, start date, end date, credits used, and session duration.
The Reservations page shows any currently running GPU kernels, the credits assigned to the reservation session, and the reservation start and end times.
Anaconda offers credits to access GPU compute time. One credit equals 100 real-world seconds. Note that this is different from how high-compute seconds are calculated for CPU use.
Credits are automatically deducted as reserved time is used. For example, if you use 10 minutes of reserved time, 6.00 credits will be deducted from your account.
View your remaining credits on the Usage Overview page.
At this time, Anaconda is offering 500 free credits to select paid subscribers only. See note for details on upgrading your subscription to gain access.
Reload the Anaconda Notebooks page and then click “Launch a Remote Kernel” from the Launcher. If you’re still unable to run a GPU kernel, file a support ticket.
