Improving GPU Robustness by making use of faulty parts

With hundreds of processing units in current state-of-the-art graphics processing units (GPUs), the probability that one or more processing units fail due to permanent faults, during fabrication or post deployment, increases drastically. In our experiments we found that the loss of a single streaming multiprocessor (SM) in an 8-SM GPU resulted in as much as 16%performance loss. The default method for dealing with faulty SMs is to turn them off. Although faulty SMs cannot be trusted to completely execute a single kernel (program assigned to an SM) correctly, we show that we can still make use of these SMs to improve system throughput by generating and supplying high-level hints to other functional SMs. By making the faulty SMs supply hints to functional SMs, we have been able to achieve an average speed-up of about 16% over the baseline case (wherein the faulty SMs are turned off). The proposed technique requires minimal hardware overhead and is highly scalable.