3D/ML Software Engineer at AMD - using fastai and nbdev!

Our team at Amd is hiring a 3D/ML software engineer. We’re a diverse team making rapid progress on an extremely important, real-time 3D graphics challenge using fastai, nbdev, pytorch and julia.

3D/ML Software Engineer - 93183 (amd.com)

We use fastai every day for all of our training on a custom, very large 3D dataset. fastai was a huge timesaver! Thanks a lot to Jeremy, Sylvain and everyone else who contributed! nbdev also works great although it took a while to figure out the optimal git workflow for our team. (We’re now around 10 full time ML developers on this project, with many more about to join/pitch in from other teams.)

AMD HR modified my original job posting to fit in their template, here are my personal remarks (not AMD’s!):

I don’t care whether you have a university degree as long as you can show you’ve built reasonably complex and original 3D or ML stuff. We don’t care what you look like, sound like, and so on – as long as you’re not a jerk, and you focus on getting the right things done.

This position is for any metropolitan area where AMD has an office, e.g. Toronto, California, Boston, Texas, etc. Long-term remote is also an option in many regions of the world.

I won’t be able to answer all questions publicly, but I can try to answer some. Thanks for your attention!

BTW I answered a few more questions in the fastai discord, in the chitchat channel.

8 Likes

Hi @claforte,

This job sounds super interesting. It’s awesome seeing fast.ai and nbdev being used so heavily at AMD. I had a couple of questions:

  1. You mentioned in the fast.ai discord that you’d like to reimplement the training code in Julia and Flux to experiment with smaller quantized types that future GPUs will be optimized for. How much does the work this team does influence the capabilities of future HW that AMD makes?
  2. You also mentioned that a key challenge is optimizing inference time so that the model can run real-time in games. Would the problem be easy if you were able to use use a large, slow model? Or is the problem hard even with a large model, and so creating a model that both performs well enough and is fast enough will be even harder?
  3. Is data collection/labeling part of the job? Or (since the dataset is 3D graphics) are the data and labels created programmatically, so there’s no need to label them by hand?
  4. What’s the breakdown between research vs. production in this role? Will it be mainly research-focused initially and then become production-focused once a viable solution is developed?

Thanks! I’m excited to send in my application!

I am also curious to know if this role can be remote, I am based in EU.