DRIFT: Harnessing Inherent Fault Tolerance for Efficient and Reliable Diffusion Model Inference
2026-04-10 • Hardware Architecture
Hardware Architecture
AI summaryⓘ
The authors look at ways to make diffusion models, which are used for creating images, run faster and use less energy. They created a system called DRIFT that carefully adjusts the computer's voltage and speed to save power or speed up the process without hurting the image quality. They also made a method to fix important errors by going back a few steps in the process when needed. Their tests show DRIFT can save a lot of energy or make inference much faster while keeping the output good.
diffusion modelDVFSfault tolerancevoltage underscalingoverclockingrollback algorithmresilience analysisenergy savingsinference latencyaccelerators
Authors
Jinqi Wen, Tong Xie, Runsheng Wang, Meng Li
Abstract
Diffusion model deployment has been suffering from high energy consumption and inference latency despite its superior performance in visual generation tasks. Dynamic voltage and frequency scaling (DVFS) offers a promising solution to exploit the potential of the underlying accelerators. However, existing approaches often lead to either limited efficiency gains or degraded output quality because they overlook the inherent fault tolerance of the diffusion model. Therefore, in this paper, we propose DRIFT, a novel algorithmarchitecture co-optimization framework that harnesses the fault tolerance for efficient and reliable diffusion model inference. We first perform a comprehensive resilience analysis on representative diffusion models. Building on these observations, we introduce a fine-grained, resilience-aware DVFS strategy that selectively protects error-sensitive network blocks and timesteps, and a rollback algorithm-based fault tolerance (ABFT) mechanism that adaptively corrects only critical errors by reverting to previous timesteps. We further optimize offloading intervals and reorganize data layouts to reduce memory overhead. Experiments across diverse models and datasets show that DRIFT can achieve on average 36% energy savings through voltage underscaling or 1.7x speedup via overclocking while maintaining generation quality.