For automotive OEMs and Tier-1 suppliers, developing production-grade perception models presents a significant bottleneck: achieving the accuracy required for safety-critical applications such as forward collision warning and autonomous emergency braking demands systematic hyperparameter optimization (HPO). However, HPO requires training hundreds of model variants, each taking hours on a single GPU, making thorough exploration of the parameter space prohibitively slow and expensive with conventional single-GPU setups.
We aim to address this challenge as part of an ongoing collaboration between AMD Silo AI and the Autoware Foundation, announced in December 2025 (for more details on the collaboration, see Advancing Open Autonomous Driving Perception at Scale). As part of this effort, we are accelerating Autoware’s end-to-end AI model training and deployment pipeline using AMD Instinct™ GPUs and ROCm™ software, uniting AMD’s scalable compute platform with Autoware’s open-source autonomous driving ecosystem - with a focus on optimizing Autoware’s perception models for production-grade Level 2+ autonomy targeting automotive OEMs, ensuring seamless performance from training to real-time inference.
By default, a single AMD Instinct™ MI300X exposes just 8 GPU devices, limiting parallelism. With GPU partitioning, the same AMD Instinct™ MI300X can be split into many more logical GPUs (up to 64), allowing a much higher number of parallel HPO workloads to run on a single node. In our experiments here, we show that GPU partitioning can accelerate HPO — all without requiring any changes to user code or additional programming effort — making high-quality, systematic optimization more economically viable for production AI pipelines.