1 of 9

XPU Dilemma

by Akshit, Etienne, Raymond, and Pablo

2 of 9

The Intel Technology

  • The new Intel technology allowed for deployment of fine tuned models on the cloud which could be trained and used on their XPU Technology
  • XPU allows for the best hardware use for the specific task
  • X being a sort of variable that can be filled by the type of task it is.

​

3 of 9

Why We Wanted To Use It

  • Offload expensive tasks to the cloud
  • Take advantage of the smart “X” and have it use the best hardware for our task
  • Use the Image to Image model, and tune it further than already provided to fit our use case.

​

  • Learn/Improve our skills with linux and cloud services.

4 of 9

The Issue

  • The hardware instances available all had us connect through a vm.
  • This vm did not have direct access to the XPU and therefore could not take advantage of all of the intel optimization packages

​

5 of 9

The “Workaround” - Theoretical Solution

  • GPU! Instead of the XPU, we know we need a GPU so we could just go straight to it!
  • API to use the Image to Image models, send image and prompt, receive images
  • NGROK to forward the port of the API to public so we can take advantage of the cloud server provided by Intel

6 of 9

The Issues Continued…

  • No access to the GPU
  • CPU too slow to use for Image to Image generation (20-30 minutes per image)
  • Timeouts before we can get Data back due to slow CPU
  • Only way to generate the image would be through the CPU which is unfeasible

​

7 of 9

lscpu - shows only access to cpu

  • ​

8 of 9

What We Got Working

  • Image To Image Generation
  • CPU Image Generation
  • Flask API Server
  • NGROK Port Forward

​

The Reason it did not work

  • CPU Too slow to work with, XPU not available.

9 of 9

Prompt to Image: Blue Elephant

with picture of elephant

​

Image to Image: Prompt: Fan with picture of fan