https://pearl-lab-vt.github.io/ PEARL builds the systems software that brings demanding AI within reach of the hardware people already have, from wearables and edge devices to warehouse-scale machines. en-us Sat, 01 Aug 2026 15:51:29 +0000 https://pearl-lab-vt.github.io/news/pearl-website-launch/ https://pearl-lab-vt.github.io/news/pearl-website-launch/ Sat, 01 Aug 2026 12:00:00 +0000 The lab now has a home at pearl-lab-vt.github.io. PEARL works on what AI costs to run: the runtimes, memory systems and schedulers that decide whether a model needs a rack of accelerators or fits on a wearable. The site collects our research, our people and our publications, and it is where lab news will appear first. news https://pearl-lab-vt.github.io/news/eccv26-diffpro/ https://pearl-lab-vt.github.io/news/eccv26-diffpro/ Mon, 06 Jul 2026 12:00:00 +0000 DiffPro sets numerical precision jointly across diffusion timesteps and network layers rather than fixing it once for the whole model. The paper appears at the European Conference on Computer Vision in Malmö this September. Congratulations to Farhana Amin, Sabrina Afroz, Kanchon Gharami and Mona Moghadampanah. The work grew out of the SC'25 posters on hybrid parallel diffusion and memory-aware coarse-to-fine inference. papers https://pearl-lab-vt.github.io/news/ismm26-gpu-prefetching/ https://pearl-lab-vt.github.io/news/ismm26-gpu-prefetching/ Tue, 16 Jun 2026 12:00:00 +0000 Unified memory makes GPUs easier to program and slower to run. Yunqi Shen presented our recovery of that lost performance at the ACM SIGPLAN International Symposium on Memory Management in Boulder: a prefetcher that predicts access patterns with a Markov model instead of reacting to page faults once the stall has already happened. papers https://pearl-lab-vt.github.io/news/yuze-li-phd/ https://pearl-lab-vt.github.io/news/yuze-li-phd/ Fri, 15 May 2026 12:00:00 +0000 Congratulations to Dr. Yuze Li, who defended his dissertation on cross-stack improvement of memory efficiency, co-advised with Ali R. Butt. The work runs from the language virtual machine down through tiered memory to the operating system, and each layer needed the others to make sense. people https://pearl-lab-vt.github.io/news/katelyn-crumpacker-msc/ https://pearl-lab-vt.github.io/news/katelyn-crumpacker-msc/ Fri, 15 May 2026 12:00:00 +0000 Congratulations to Katelyn Crumpacker on defending her M.Sc. thesis, "LLM-Guided Runtime Parameter Optimization for Energy-Efficient Model Inference." The runtime knobs that decide what inference costs in energy are numerous, interacting, and rarely tuned. Her work uses a language model to search that space. people https://pearl-lab-vt.github.io/news/sony-scale-r/ https://pearl-lab-vt.github.io/news/sony-scale-r/ Mon, 04 May 2026 12:00:00 +0000 Sony has funded SCALE-R, on semantic co-rendering and adaptive link-aware encoding for real-time graphics. The award runs to May 2027 and continues a collaboration on efficient inference and media pipelines at the edge that began with WASML, our 2023 Sony Faculty Innovation Award on scalable serverless computer vision. grants https://pearl-lab-vt.github.io/news/fact-preprint/ https://pearl-lab-vt.github.io/news/fact-preprint/ Wed, 29 Apr 2026 12:00:00 +0000 Ask a language model for a fast CUDA kernel and it will hand back plausible code that often does not compile, let alone run quickly. FACT grounds the synthesis in CUTLASS C++ across a three-stage agentic workflow. Auto-tuned kernels beat PyTorch cuBLAS by 1.06x to 1.18x on GEMM workloads, and composing fused multi-head attention with a fused MLP GEMM+GELU reaches a 2.79x end-to-end speedup on a MiniGPT block. Preprint by Sina Heidari. papers https://pearl-lab-vt.github.io/news/llm-runtime-params-preprint/ https://pearl-lab-vt.github.io/news/llm-runtime-params-preprint/ Wed, 29 Apr 2026 12:00:00 +0000 The runtime parameters that decide what inference costs in energy are numerous and they interact, so in practice nobody tunes them by hand. Katelyn Crumpacker's framework puts a chat-based LLM in the loop with structured feedback prompts. The enhanced prompt template converges in 3.4 prompts on average against 5.2 for the baseline, beats Sobol sampling on convergence speed, and reaches lower final energy per token across hardware setups. papers https://pearl-lab-vt.github.io/news/ics26-dllm-serve/ https://pearl-lab-vt.github.io/news/ics26-dllm-serve/ Tue, 28 Apr 2026 12:00:00 +0000 Diffusion language models generate text by iterative refinement, and the memory traffic that requires has kept them out of production serving. dLLM-Serve closes that gap. The paper appears at the 40th ACM International Conference on Supercomputing in Belfast. Congratulations to Jiakun Fan, Yiyuan Zhang and Xiangchen Li. papers https://pearl-lab-vt.github.io/news/sweet-frontiers-2026/ https://pearl-lab-vt.github.io/news/sweet-frontiers-2026/ Wed, 15 Apr 2026 12:00:00 +0000 SWEET partitions and quantizes a model across devices so that no stage of an edge inference pipeline becomes the bottleneck for the rest. The journal version appears in Frontiers in Complex Systems, closing a major thread of the NSF SWEET project with Hans Vandierendonck and Deepu John. papers https://pearl-lab-vt.github.io/news/sigmetrics26-wisp/ https://pearl-lab-vt.github.io/news/sigmetrics26-wisp/ Wed, 25 Mar 2026 12:00:00 +0000 Speculative decoding wastes work whenever the draft model guesses wrong, and at the edge that waste competes with the very requests it was meant to accelerate. WISP suppresses both through dynamic drafting and SLO-aware batching. The paper appears in Proceedings of the ACM on Measurement and Analysis of Computing Systems, led by Xiangchen Li and Jiakun Fan with collaborators at Queen's University Belfast, University College Dublin and Virginia Tech. papers https://pearl-lab-vt.github.io/news/capwic26-marco/ https://pearl-lab-vt.github.io/news/capwic26-marco/ Sat, 14 Mar 2026 12:00:00 +0000 MARCO puts several coordinated agents to work optimizing high-performance code, and was presented at the ACM Capital Region Celebration of Women in Computing. The paper comes from our undergraduate research cohort working with Sina Heidari and Margaret Ellis, and is the first published result of the lab's agentic-AI-for-HPC effort. Authors: Ariana Nafisi, Veljko Cvetkovic, Claire Shin, Kathleen Reece, Krishnateja Reddy, Bryan Torres, Sina Heidari, Margaret Ellis and Dimitrios S. Nikolopoulos. papers https://pearl-lab-vt.github.io/news/iwapt26-modality-inflation/ https://pearl-lab-vt.github.io/news/iwapt26-modality-inflation/ Mon, 02 Mar 2026 12:00:00 +0000 Adding a second modality to a language model inflates its energy cost far beyond the extra tokens involved. Mona Moghadampanah, Adib Rezaei Shahmirzadi and Farhana Amin measure where that cost comes from and where it can be recovered. To appear at the 21st International Workshop on Automatic Performance Tuning in New Orleans. papers https://pearl-lab-vt.github.io/news/ipdps26-apex/ https://pearl-lab-vt.github.io/news/ipdps26-apex/ Fri, 20 Feb 2026 12:00:00 +0000 APEX runs online LLM inference on GPUs too small to hold the model, overlapping CPU and GPU execution so neither waits on the other. The paper appears at the 40th IEEE International Parallel and Distributed Processing Symposium in New Orleans. Congratulations to Jiakun Fan, Yiyuan Zhang and Xiangchen Li. papers https://pearl-lab-vt.github.io/news/wacv26-polymorph/ https://pearl-lab-vt.github.io/news/wacv26-polymorph/ Tue, 10 Feb 2026 12:00:00 +0000 Polymorph performs multi-label classification on video streams within an embedded power budget, presented at the IEEE/CVF Winter Conference on Applications of Computer Vision in Tucson. Joint work with Saeid Ghafouri, Mohsen Fayyaz, Xiangchen Li, Deepu John, Bo Ji and Hans Vandierendonck. papers https://pearl-lab-vt.github.io/news/tdis26-configspec/ https://pearl-lab-vt.github.io/news/tdis26-configspec/ Mon, 26 Jan 2026 12:00:00 +0000 Distributed speculative serving offers more configurations than anyone can tune by hand: which draft model, which split across edge and cloud, which batching policy. ConfigSpec selects among them from profiles. To appear at the 4th International Workshop on Testing Distributed Internet of Things Systems in Edinburgh. papers https://pearl-lab-vt.github.io/news/emad-abdrabou-phd/ https://pearl-lab-vt.github.io/news/emad-abdrabou-phd/ Mon, 15 Dec 2025 12:00:00 +0000 Congratulations to Dr. Emadeldin Abdrabou, who defended his dissertation on optimizing edge server scaling and placement for Open Radio Access Network slicing, co-advised with Scott Midkiff. He has taken up a position as Assistant Professor at Sohag University. people https://pearl-lab-vt.github.io/news/sec25-sled/ https://pearl-lab-vt.github.io/news/sec25-sled/ Thu, 04 Dec 2025 12:00:00 +0000 SLED brings speculative decoding to edge serving, where the draft model and the target model may sit on different machines with a slow link between them. Presented at the Tenth ACM/IEEE Symposium on Edge Computing in Washington, D.C., led by Xiangchen Li with Dimitris Spatharakis, Saeid Ghafouri and Jiakun Fan, and collaborators at Queen's University Belfast and University College Dublin. papers https://pearl-lab-vt.github.io/news/sc25-posters/ https://pearl-lab-vt.github.io/news/sc25-posters/ Tue, 18 Nov 2025 12:00:00 +0000 The lab presented four posters at SC'25 in St. Louis, covering long-context prefill parallelization, hybrid parallel diffusion, precision optimization for diffusion inference, and energy characterization of multimodal LLM inference. A strong showing from a first-year Ph.D. cohort. HydraCache: Long-Context Prefill Parallelization via Distributed Cache Blending — Adib Rezaei Shahmirzadi, Sadegh Shabihi, Feiyang HuangDivide, Conquer, and Denoise: Hybrid Parallel Diffusion with Memory-Aware Coarse-to-Fine Inference — Farhana Amin, Kanchon GharamiDiffPro: Joint Timestep and Layer-Wise Precision Optimization for Efficient Diffusion Inference — Farhana Amin, Kanchon GharamiEnergy-Efficient Multimodal LLM Inference: Stage-Level Characterization and Input-Aware Controls — Mona Moghadampanah, Adib Rezaei Shahmirzadi papers https://pearl-lab-vt.github.io/news/vmil25-memory-tiering/ https://pearl-lab-vt.github.io/news/vmil25-memory-tiering/ Tue, 14 Oct 2025 12:00:00 +0000 Python programs place data across memory tiers badly, because the interpreter knows what objects mean while the operating system knows where the pages are, and neither tells the other. Our work on memory tiering inside the virtual machine appeared at the 17th SPLASH Workshop on Virtual Machines and Language Implementations in Singapore. Congratulations to Yuze Li, Shunyu Yao and co-authors. papers