AIBooster Boosts GPU Utilization and Cost Awareness for Autonomous Driving AI Training

Masahiro Tamori:We already had a strong working relationship with Fixstars. The immediate catalyst was a conversation between our President, Izumi Kawanishi, and Fixstars’ President, Satoshi Miki. When I heard about a tool that could double the execution efficiency of our computational resources, we decided to try it right away.

A key reason for this was our heavy use of GPUs on the cloud for AI, specifically for developing recognition models. The hardware costs are extremely high, so we needed to accurately understand our return on investment and confirm if we were truly using our GPU resources efficiently. We also wanted a way to visualize and maximize that utilization.

Yasuhiro Suto: From a technical standpoint, the deciding factor was the ability to monitor the GPU's "SM Activity" (Streaming Multiprocessor Activity: the computational efficiency at the basic GPU unit level) in real time with great detail. We had been measuring per-GPU and per-unit-of-time execution efficiency ourselves, but not to the level of detail AIBooster provides. This functionality was what resonated most with our needs.

Masahiro Tamori:The ability to monitor our GPU usage in the cloud environment in real time has been a huge benefit.

Motoi Kataoka: The most significant effect has been a heightened cost awareness. President Kawanishi himself checks the dashboard, showing a strong interest in GPU utilization efficiency. This interest from the top has prompted every AI engineer and every member of the acceleration team to become more conscious of how efficiently their own tasks are using GPUs.

Compared to other monitoring tools, the dashboard is well-organized and easy to view, which is a big plus.

Yasuhiro Suto: In terms of features, the recent update that allows us to check the status of a job while it's still running is excellent. We used to only be able to check after a job was complete, but this increased real-time capability has sped up our improvement cycle considerably.

Now, whenever we implement a performance improvement measure, the first step is to analyze the current situation with AIBooster. The function to analyze both GPU and storage usage on a per-job basis is also unique and highly effective as a starting point for multi-faceted analysis.

Masahiro Tamori:From a tool perspective, we hope to see continuous improvements in usability, such as real-time result checking and the ability to interact with jobs while they are running.

We also hope the reports will go beyond just numerical data, including the reasoning behind the results and the progress of improvements, so we can perform a deeper analysis.

Regarding support, we would be grateful for your help with upstream issues beyond just acceleration and monitoring, such as "stabilizing the AI training process" and "scheduler optimization."

Yasuhiro Suto: I'd like to see AIBooster further automate the optimization of AI processing.^* The ideal scenario is that AIBooster automatically handles about 80% of the work, allowing engineers to focus on the remaining 20% of core tasks.

As our training data grows, the losses from rare training job failures are not insignificant. If AIBooster could detect the anomalous spikes that cause these failures, we could take measures to prevent them, which would increase the stability of our training process.

Masahiro Tamori: Given the current GPU shortage, we hope for advanced technical support that would allow us to combine different types of GPUs and achieve performance equivalent to a single high-end GPU.

*The contents of this article are based on information from the time of the interview (September 2025). Titles and affiliations are as of the article's publication.