Facts, Issues and Questions - GPUs for Dependability

Abstract

Listen

Translate article

Graphics Processing Units (GPUs) offer massive parallelism, comprising many actual paradigms like manycore, multithreading and SIMD. Today, nearly every computer is equipped with at least one graphics card, containing one or more GPUs bringing massive parallelism to the desktop. GPUs are usually used in their main function, that is, to compute visibility, lightning, perspective, etc. in games. As this technology is widely used, it is lowcost. In the majority of the cases, graphic cards do not spend their entire lives by executing game code. Thus, such a massive parallel system is underchallenged most of the time. Shortly after the availability of comfortable programming environments, based on CUDA (Compute Unified Device Architecture) or HLSL (high-level shader language), researchers have become interested in using this power for general-purpose computing (GPGPU, General-Purpose computing on the GPU). Thus, different applications originated, e.g. physics, cryptography 0, DNA sequencing 0 and medical imaging. For further examples and overview, see 0 and 0. The trend to compute such workloads with GPUs will go on as the DirectX 11 (compute) or the OpenCL 0 standards show. The fault-tolerant execution of (sensible) workloads on GPUs was – to the knowledge of the author – never proposed. Sensible computations should be carried out in a reliable way. What is the sense of a computation to find a private key if the program is correct but the hardware is subjected to faults and the program never finds the key? E.g. transient faults can be caused from fluctuations in the main current, radiation or RAMs not running within their specification etc. What if an encryption is faulty due to temporal faults or how can we detect a faulty medical diagnosis? The need to do computations precisely has led to the development of more sophisticated and sometimes expensive graphics processing units 0, needed by CAD applications. Larrabee 0 is a manycore visual computing architecture. It uses multiple in-order x86 CPU cores that are augmented by a wide vector processor unit, as well as some fixed function logic blocks. This provides much higher performance per watt and per unit of area than out-of-order CPUs on highly parallel workloads. Vision4ce 0 launched a new line of General-purpose Rugged Image Processing (GRIP) products at the recent SPIE Defense and Security Symposium. The GRIP-Beta showed GPGPU-based image processing demonstrations, analog and Gigabit 4