Cross-Layer Adaptive Hardening Strategies for Next-Generation General-Purpose Graphics-Processing Units (GPGPUs)

• Applicant: Professor Dr.-Ing. Michael Hübner, since 9/2021
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems ; Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 458578717

Graphics Processing Units (GPUs) are currently being employed in a broad range of domains where the long-term dependable and error-free execution is of uttermost importance, such as in automotive, avionics, and space domains. Initially designed for multimedia processing, however, and since multimedia remains the largest market segment, the microarchitecture of modern GPUs is still not well-suited for dependable execution. In domains where that constraint is of uttermost importance, GPU designs are extended only by protecting the memory arrays (which are easily replaceable without changing the entire design) with Error Correction Codes (ECC). Additional measures for fault-tolerant execution, required to achieve the strict low error requirements, are then implemented at the software level. This scenario demonstrates a clear need for microarchitectures that can serve different application domains while keeping the design costs as low as possible. To make matters even more challenging, all these applications present strict performance, power and energy consumption constraints which must also be met. In order to address the challenging and conflicting requirements of generality (having a device that serves multiple domains) and efficiency (performance, energy consumption, and dependability), this research proposes the investigation of novel adaptive GPU architectures that are capable of dynamically optimizing the execution towards domain/application-specific constraints such as performance, low energy consumption, and dependable execution by adapting to environmental conditions (such as the probability of faults) and application characteristics (such as the resource utilization). Moreover, to achieve the strict reliability levels required by future safety-critical systems such as those of the ISO 26262 standard, reliability will be addressed from a cross-layer perspective, employing techniques that range from the application to the hardware level in order to achieve the dependability target while studying the combination of approaches that provides the best tradeoff with respect to performance and energy consumption.This project will result in a novel design methodology for adaptive and dependable GPUs, with new cross-layer and adaptive strategies that enable these devices to fit multiple domains, optimizing at run time to the instantaneous requirements of each application (high performance, high energy efficiency, or high dependability). Below, we motivate the importance of GPUs and the state-of-the-art in design for dependability and GPU-specific approaches to that, and the previous work by the proponent in the fields of adaptability, reliability, and GPUs.

• DFG Programme: Research Grants
• International Connection: Brazil
• Cooperation Partners: Professor Dr. Antonio Carlos Schneider Beck; Professor Dr. Paolo Rech
• Ehemaliger Antragsteller: Dr. Marcelo Brandalero, until 9/2021