Abstract: In recent years, there has been growing interest in executing Artificial Intelligence algorithms directly onboard satellites. AI can significantly optimize data storage and transmission, allowing real-time payload processing, minimizing latency, bandwidth usage, and overall operational costs. Despite the advantages, however, Machine Learning algorithms are computationally demanding. Thus, one of the main challenges lies in developing hardware accelerators that not only deliver high computational performance but also endure the harsh conditions of the space environment. Hardware architectures like Graphics Processing Units are typically used either for cloud applications due to their high power consumption or for short-duration missions in Low Earth Orbit due to reliability concerns. GPU@SAT, a hardware/software framework provided by IngeniArs S.r.l., represents a compelling solution for executing AI tasks in space. By integrating a Soft GPU IP core, it harnesses the parallel processing power of General-Purpose Computing on GPUs while leveraging the Space-Qualified hardware of Field Programmable Gate Arrays. This work focuses on extending the Instruction Set Architecture of GPU@SAT to support vector operations, improving compatibility with ML frameworks that use 8-bit quantization, such as TensorFlow Lite. We present implementation results on the Radiation-Tolerant Xilinx XQRKU060 FPGA, along with performance benchmarks demonstrating speedups and energy efficiency improvements enabled by the vector extension.

Keywords: Vector Extension; Soft GPU; Artificial Intelligence; Machine Learning; Hardware Acceleration; Space