1st SAFARI Live Workshop (SaLWo) on Storage Systems

1st SAFARI Live Workshop (SaLWo) on Storage Systems

🎙 Onur Mutlu Lectures 👥 64K 📅 September 19, 2025 ⏱ 500 min 👁 922 📄 original study 🧭 2026-08-16
Available in: English (current) Français

Keywords

RISEin-storage processingapproximate nearest neighbor searchretrieval-augmented generationflash memory

Summary

This video is a recording of the first SAFARI Live Workshop (SaLWo) on Storage Systems, held on September 19, 2025, at ETH Zurich. The workshop focuses on the latest advances in storage systems, with a primary presentation by Andreas Kosmas Kakolyris on RISE, a retrieval system with in-storage processing. RISE addresses the data movement bottlenecks in retrieval-augmented generation (RAG) pipelines by accelerating both nearest neighbor search and document retrieval within the storage system. The presentation covers the background of RAG, the limitations of existing approaches, the design of RISE, and its evaluation. Key results show that RISE improves throughput by 15x and energy efficiency by 55x over CPU-based systems, reduces IO data movement to under 0.2% of total RAG runtime, and outperforms state-of-the-art in-storage processing accelerators by up to 2.5x. The talk includes a detailed Q&A session where the presenter clarifies technical aspects, such as the use of existing flash logic for distance computation and the handling of data placement. The video is highly technical and suitable for researchers and practitioners in computer architecture and storage systems.

177 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides substantial value by presenting a novel system (RISE) that tackles a critical bottleneck in RAG pipelines. The argumentation is solid: the presenter identifies a clear problem (data movement overheads), analyzes existing solutions (quantization and in-storage processing) and their limitations, and then proposes a comprehensive solution with three key mechanisms. The evaluation is thorough, comparing RISE against multiple baselines (CPU, no-IO, and state-of-the-art ISP accelerator) across various datasets and SSD configurations. The results are compelling, showing significant improvements in throughput, energy efficiency, and end-to-end latency. The Q&A session further strengthens the argumentation by addressing potential concerns about implementation details and scalability. Overall, the video offers a well-structured and evidence-based presentation of original research.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high, as the work is published at ISCA 2025, a top-tier venue in computer architecture. The presentation includes detailed methodology and evaluation, and the description provides links to the paper and related publications. The sources cited are relevant and from reputable venues (e.g., ISCA, ASPLOS, IEEE TCAD). The title accurately reflects the content, which is a workshop on storage systems. The video is a live recording, so there are minor production imperfections, but they do not detract from the technical content. The description also includes links to recommended readings and lecture playlists, which are useful for further study. Overall, the sources are credible and the title-content alignment is excellent.

243 words

Title / Content Match

The title accurately describes the content: a live workshop on storage systems, specifically the fourth day of the SAFARI Live Workshop.

Quality & Reliability

9/10

The video is a live workshop recording from a reputable academic group (SAFARI, ETH Zurich) presenting original research (RISE) published at ISCA 2025. The presentation includes detailed technical content, methodology, and evaluation results. The description provides links to the paper and related publications, enhancing credibility. Minor limitations include the informal setting and lack of peer review in the video itself, but the underlying work is peer-reviewed.

Key Moments

Cited Sources

Concurring Sources

External References

Contribution & Novelties

The video presents RISE, a novel retrieval system with in-storage processing that accelerates both nearest neighbor search and document retrieval in RAG pipelines. The key innovation is leveraging existing computational resources within SSDs (e.g., page buffer logic and ECC circuitry) to perform distance calculations and filtering, eliminating the need for significant hardware modifications. This approach reduces data movement overheads dramatically, achieving 15x throughput and 55x energy efficiency improvements over CPU-based systems. The work also introduces a lightweight embedding-to-document linkage mechanism using the out-of-band region of flash pages, and an ISP-tailored data placement algorithm. These contributions advance the field of storage-centric computing and have practical implications for RAG-based AI systems.

Pour aller plus loin :

152 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a technically deep, well-sourced, and highly informative presentation. The video excels in providing detailed quantitative results and rigorous evaluation, with a strong focus on practical implementation. The only slight weakness is the informal setting, but this does not detract from the overall quality.

Reliability 9/10

💬 Sur les 0 commentaires analysés, aucune tendance n'est disponible.