Can Driverless AI on GCP make use of K80 GPU for improved performance?

Other Questions ai employee
1

I’m testing how well the H2O Driverless AI instance performs on Google Cloud Platform. I added a K80 GPU to the instance, but when I run my experiments with ‘GPU enabled’, it seems like the GPU isn’t being utilized. Are there specific installation or configuration steps I need to follow for H2O to work with the GPU?

2

had this exact problem last month. K80s are basically dead for DAI now - can’t even force them to work anymore. I tried downgrading to older DAI versions but hit other compatibility issues. switched to a T4 instance and everything worked right away. way faster too, even tho it’s single GPU vs the K80’s dual setup.

3

The K80 GPU is quite outdated, and H2O Driverless AI has phased out support for older GPUs in recent versions. I encountered this issue with legacy hardware last year as well. The K80 operates on the Kepler architecture, which does not work well with the newer CUDA versions that Driverless AI requires. To effectively utilize your GPU with the latest DAI versions, you’ll need a GPU based on at least the Pascal architecture, such as the GTX 1060/1070 or Tesla P4/P40. Check the DAI documentation for the minimum CUDA compute capability requirements, which are generally 6.0 or higher now, while the K80 only supports 3.7. If GPU acceleration is essential, consider upgrading to a T4 or V100 instance.

4

Yeah, JackWolf69’s right about K80 compatibility - total headache.

But here’s what I’d do instead: automate your ML pipeline setup. Skip the hardware wrestling entirely.

I built workflows that auto-provision GCP instances with compatible GPUs based on what I need. No more wondering if hardware will work.

My setup:

  • Spins up T4/V100 instances for GPU work
  • Drops to CPU-only for light tasks
  • Handles CUDA and driver installs automatically
  • Monitors usage and switches configs as needed

Saved me tons of debugging time. Rather than fixing your K80, automate the whole thing to use proper hardware from day one.

Latenode makes this infrastructure automation pretty easy. Set triggers based on your experiments and let it handle everything.