The Future of SCIENTIFIC Computing is Interactive

This demonstration showcases a breakthrough in real-time, interactive high-performance computing using a popular video game as a test case. Despite technical and administrative challenges, the experiment successfully implemented:

A custom-built compute cluster.

Real-time interaction with an AI agent.

Simultaneous training and gameplay.

This achievement addresses previously unresolved issues in interactive AI training and gameplay, opening new possibilities for high-performance computing applications.

This experiment involves StarCraft II, a popular video game that requires admin privileges to install non-academic software on the machine clusters at the academy. Instead, I had to provision my own cluster. Training the bot demands modest GPU power; the small-scale replication setup used 56 NVIDIA A100 GPUs in previous research work, but interestingly, running these for 2 days was sufficient. I had an idea in mind that differs from typical batch jobs, and I have made things interactive.

My real-time setup and solution already address those challenges:

While there will always be accessible supercomputer clusters from academia or through research grants or cloud credits, notably, interactive playing and simultaneous training with the agent were thought to be impossible and unresolved*. However, by exploring a different network protocol, I managed to resolve graphics software issues, and hence running the program now creates two instances of the game: human vs. bot or bot vs. bot, and allows real-time play.

Another demo of this is real-time rendering which is critical for either a virtual reality or mixed reality future.

What it seems to me is that we can embark on an adventure, unlock compute, and create a future where high-performance computing is not just a frozen AI model output of a batch command, but can be experienced in real-time with experiential training in our quest for 42, and for more information on my research papers related to this project, please visit the following links, e.g., scavenging network protocols, AI safety in generative models and with more to follow for real-time inferencing.


Regards,
Jaymarieeexplore.ieee.org/document/7158220

ONE-Liner

Polar And Polar is a research and development company specializing in super low latency and real-time interaction with high-performance compute.

We understand that your needs are unique. Let's connect you with our professional services team for a personalized consultation. Head on over to our professional services page to connect with our team.