Gravitron Support - Price and Performance improvements

[larsll]

Describe the feature you'd like

Migration from ml.c5.4xlarge instance which is expensive and underpowered to a ml.c8g.4xlarge, which has lower instance costs and double the vCPUs.

Additional context

• The Simulation application on ROS2 Jazzy is ARM compatible with only minor adjustments in the build steps; source code requires no changes.
• The current Tensorflow configuration is limited to 1 thread which significantly slows down the policy update phase. More vCPUs would make it possible to increase the Tensorflow threads and performance, leading to more effective training (more training steps per hour).
• The increased performance should also make it possible to fix the video generation performance issues.

[larsll]

Further testing, related also to #19, shows that it should be possible to run the simulation on an ml.c8g.2xlarge as the c8g family has 8 real cores, not 4 which the Intel based 2xlarge instances have. On a 2xlarge clamping of real-time factor is an option to get more stable performance.

[abhirpat]

Thank you Lars for your suggestion. We will review this and get back to you.

[larsll]

Looking at the documentation around quotas it does not seem possible to use Gravitron (c7g, c8g) instances for Training. (There is no quota to request for training jobs, only for endpoint usage...)

[sanjay-reddy-kandi]

Thanks for raising this! We're currently following up with our internal team regarding Graviton (c7g, c8g) instance support for Training. We'll update this issue once we have more information to share.