Hello,
At our company, in addition to our traditional server rooms, we have constructed micro data centers that allow for the easy deployment of AI servers and GPU-based quantum computing servers. AI and quantum computing servers can be noisy and difficult to manage due to heat dissipation issues. By situating high-performance servers in separate units, we can utilize them without sacrificing the comfort of workspaces.
Regional and Corporate Quantum Computing & AI Micro Data Center
https://blueqat.com/yuichiro_minato2/cf66d00b-3fe3-4663-b96e-ad380b3e74e0
Furthermore, AI and quantum computing typically involve lengthy computations rather than frequent access, and the input/output data size is relatively small. Given this, we explored the use of 5G cellular networks for operating AI and quantum computing, which has proven to be stable.
5G Cellular Network-Based Generative AI and Quantum Computing Simulation Service
https://blueqat.com/yuichiro_minato2/7f79ee01-2821-446e-922c-e58bd68b7952
We have further expanded this by leveraging satellite-based internet services. Specifically, we utilized Starlink, a low earth orbit satellite at an altitude of 550 km, to directly connect to the internet for AI server and GPU quantum computing operations.
Starlink
By setting up a Starlink antenna, it automatically locates low earth orbit satellites and connects to the internet. Installation was straightforward, and the connection was established quickly. The cost is reasonable, with the initial expense for the antenna being a few tens of thousands of yen and a monthly fee of 6,600 yen for individuals or 9,600 yen for businesses.
Using this connection, I aimed to operate our servers. However, I encountered an issue: unlike with a cellular network where a global IP address can be obtained and access to the server can be managed via DDNS, I could not immediately establish a connection to issue commands for AI or quantum computations.
Solution: Websocket Implementation
Given the difficulty in directly accessing the computation servers from the outside, I implemented a polling mechanism where the computation servers periodically check in with our main server to see if there are any pending tasks. However, this approach was cumbersome due to the increased response load. Instead, I utilized websockets for a more efficient solution.
Here’s how it works:
1. A user makes a request to our server room (e.g., "What is the preparation time for this recipe?").
2. Our server sends this request via websocket through Starlink to the computation server.
3. The computation server, using the RAG (Retrieval-Augmented Generation) system, retrieves information from a pre-registered database and generates a response.
4. This response is sent back to our server room through the websocket connection and then to the user.
For example, the query "What is the preparation time for this recipe?" resulted in the response "The preparation time is approximately 30 minutes." The entire process was seamless, and using Starlink for this complex interaction felt no different from using a regular internet connection.
Managed RAG System for Enterprises
https://blueqat.com/yuichiro_minato2/1099fdd0-0bfd-4191-8821-33b488136c49
Additionally, I tested the simulation of quantum computing calculations via this setup. Although lengthy calculations might require alternative methods to obtain results efficiently, short computations were executed in real-time without issues. This suggests that actual quantum computers, which often have shorter computation times, could also benefit from this setup.
In summary, the trial was simple and brief, but we successfully executed generative AI (RAG) and quantum computing tasks using low earth orbit satellites, demonstrating performance comparable to traditional internet usage. The ability to utilize Starlink even in remote areas like Antarctica or deep jungles marks a significant advancement.
Thank you.