Context: Reliability is a mandatory requirement for multimodal AI deployed in safety- and mission-critical embedded systems, such as those used in autonomous vehicles, aerospace, and medical applications. However, hardware faults induced by external factors, such as temperature variations or radiation, can cause errors that propagate through the hardware and software stack and produce incorrect inferences. Commercial AI accelerators offer high computational efficiency but limited fault protection, while hardened accelerators sacrifice performance and energy efficiency. Additionally, software-based reliability methods scale poorly to large multimodal AI models. These applications therefore require hardware accelerators that combine high performance and reliability with low power and energy consumption.

Objectives: The RAISE project aims to make multimodal AI tolerant to hardware faults without reducing performance or increasing power or energy consumption. RAISE will combine hardware and software modifications in a cross-layer methodology. The main hypothesis is that not all hardware faults affect AI inference equally. RAISE will therefore trace faults from hardware components through AI model operations to incorrect inferences, combine sensitivity metrics with hardware fault analysis, and use selective protection to develop efficient fault-mitigation methods. These methods will be validated through hardware simulations, realistic fault-injection campaigns, and radiation and temperature experiments. The core innovation of RAISE is predictive AI reliability assessment that guides efficient cross-layer protection without depending on exhaustive testing.

Partners

• TARAN key members: Fernando Fernandes dos Santos (PI), Marcello Traiola, and Angeliki Kritikakou
• DGA key members: Youri Helen