Conduit: Programmer-Transparent NDP Using Multiple Compute-Capable Resources in SSDs - HPCA 2026

Item: Conduit: Programmer-Transparent NDP Using Multiple Compute-Capable Resources in SSDs - HPCA 2026
Rating: 4
Author: PostSingularity Institute

by Rakesh Nadig

Applied Sciences & Engineering Computing & Cybersecurity UKComputer hardware UKMEmbedded systems

🎙 Rakesh Nadig 👥 64K 📅 August 26, 2026 ⏱ 18 min 👁 11 📄 original study 🧭 2026-08-26

Available in: English (current) Français

Keywords

NDPSSDin-storage processingin-flash processingprocessing using DRAM

Summary

The talk presents Conduit, a framework for programmer-transparent near-data processing (NDP) in solid-state drives (SSDs). The authors observe that modern SSDs contain multiple heterogeneous compute-capable resources: embedded cores in the controller (in-storage processing, ISP), DRAM (processing using DRAM, PUD), and flash chips (in-flash processing, IFP). Each paradigm has trade-offs: ISP is general-purpose but slow, PUD is highly parallel but requires data movement from flash, and IFP eliminates data movement but has limited operations. A case study shows that no single paradigm is best for all workloads, and naive combination can hurt performance. Conduit addresses this by dynamically offloading computations at instruction granularity to the most suitable resource, based on a cost function that considers data movement, computation latency, data dependencies, and resource contention. The framework includes a compile-time step that vectorizes code and a runtime step in the SSD controller. Evaluation using a simulator based on Samsung 980 Pro shows that Conduit outperforms prior offloading models by 1.8x on average and achieves 62% of an ideal policy’s performance, while reducing energy by 46% compared to data-movement-based offloading. Overheads are low: ~17 bytes of DRAM for cost features and <4 microseconds per instruction. The talk concludes with a Q&A discussing assumptions on IFP and extensibility to dedicated accelerators.

207 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a strong value proposition by addressing a gap in SSD-based NDP: prior works typically map computations to one or two paradigms, lacking generality and programmer transparency. Conduit’s key contribution is the holistic, dynamic offloading at instruction granularity, which is well-motivated by the case study showing workload-dependent bottlenecks. The argumentation is solid: the authors systematically analyze trade-offs, adapt offloading models from other domains, and evaluate against realistic baselines. The cost function is comprehensive, considering multiple factors. The evaluation is thorough, covering performance, energy, and overheads, with a simulator based on a real SSD. The claims are supported by quantitative results. The Q&A shows awareness of limitations and future extensions. However, the talk is a presentation, so some details are omitted, but the argumentation remains convincing.

Scientific Rigor, Source Quality, Title Accuracy

The talk is based on a peer-reviewed paper at HPCA 2026, which lends credibility. The presenter cites prior works in the field, such as Ambit, C DRAM, MIM DRAM, Flash Cosmos, and ADIS Flash, and provides references in the description. The methodology is described with enough detail to be reproducible (simulator based on MPsim, Samsung 980 Pro configuration). The title accurately reflects the content. The talk does not include any advertising or sponsored content. The description includes links to slides and recommended reading, which are relevant and credible. Overall, the scientific rigor is high, with clear sourcing and appropriate context.

242 words

Title / Content Match

The title accurately reflects the content: the talk presents Conduit, a framework for programmer-transparent near-data processing using multiple compute-capable resources in SSDs.

Quality & Reliability

8/10

Presentation of a peer-reviewed academic work (HPCA 2026) with detailed methodology, quantitative evaluation, and references to prior art. The talk is technical and specific, with clear claims supported by simulation results. Minor limitations: reliance on simulations and assumptions about IFP capabilities, but overall rigorous.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Conduit slides (PDF) — Slides of the presentation
Conduit slides (PPTX) — Slides of the presentation
A Modern Primer on Processing in Memory — Recommended reading on processing in memory
Memory-Centric Computing: Solving Computing's Memory Problem — Recommended reading on memory-centric computing
Memory-Centric Computing: Recent Advances in Processing-in-DRAM — Recommended reading on processing-in-DRAM
Intelligent Architectures for Intelligent Computing Systems — Recommended reading on intelligent architectures
RowHammer: A Retrospective — Recommended reading on RowHammer
Fundamentally Understanding and Solving RowHammer — Recommended reading on RowHammer
Accelerating Genome Analysis via Algorithm-Architecture Co-Design — Recommended reading on genome analysis acceleration
From Molecules to Genomic Variations: Accelerating Genome Analysis via Intelligent Algorithms and Architectures — Recommended reading on genome analysis acceleration

Concurring Sources

A Modern Primer on Processing in Memory — Provides foundational concepts on processing in memory, supporting the motivation for NDP.
Memory-Centric Computing: Recent Advances in Processing-in-DRAM — Discusses recent advances in processing-in-DRAM, which is one of the NDP paradigms used in Conduit.
Intelligent Architectures for Intelligent Computing Systems — Discusses intelligent architectures and computing systems, aligning with the goals of Conduit.

Dissenting Sources

No discordant sources identified — The talk does not present conflicting sources; it builds upon prior work and extends it.

External References

Contribution & Novelties

Conduit is the first general-purpose, programmer-transparent NDP framework for SSDs that dynamically offloads computations at instruction granularity to multiple heterogeneous compute resources (ISP, PUD, IFP). It addresses the limitations of prior works that map to one or two paradigms and lack generality. The framework includes a compile-time vectorization step and a runtime cost-function-based offloading mechanism that considers data movement, computation latency, dependencies, and resource contention. Evaluation shows significant performance and energy improvements over prior offloading models, with low overheads. The work is extensible to future accelerators.

Pour aller plus loin :

Processing-in-Memory (Wikipedia) — Provides background on the general concept of processing in memory.
Near-Data Processing (Wikipedia) — Explains the broader paradigm of moving computation to data.
Solid-state drive (Wikipedia) — Background on SSD architecture and components.
NVMe (Wikipedia) — Relevant to the communication interface used in the work.
RowHammer (Wikipedia) — Related to DRAM reliability issues, relevant to processing using DRAM.

152 words

Radar Profile

The radar profile shows high scores in technical level and information quality, reflecting the deep technical content and rigorous methodology. The lower score in information quantity is due to the short duration and focused scope, but the content is dense. Overall, the profile indicates a highly technical and reliable presentation.

Reliability 8/10

💬 No comments were provided for analysis.