Seminar in Computer Architecture - L2: Premiere: MIMDRAM and FC-DRAM (Fall 2026)

Seminar in Computer Architecture - L2: Premiere: MIMDRAM and FC-DRAM (Fall 2026)

by Onur Mutlu Lectures · View the channel on PSI

Applied Sciences & Engineering Computing & Cybersecurity UYComputer scienceUYFComputer architecture and logic design
🎙 Onur Mutlu Lectures 👥 66K 📅 October 1, 2026 ⏱ 95 min 👁 1K 📄 literature review 🧭 2026-10-06
Available in: English (current) Français

Keywords

processing-in-memoryDRAMMIMDcompilerenergy efficiency

Summary

The seminar presents MIMDRAM, an end-to-end processing-using-DRAM system for high-throughput, energy-efficient, and programmer-transparent multiple-instruction multiple-data (MIMD) processing. The talk begins with the problem of data movement energy and the limitations of existing processing-in-memory (PIM) architectures, particularly their coarse and rigid granularity which leads to underutilization and limited computation support. MIMDRAM addresses these issues by segmenting DRAM wordlines into smaller mats (512 elements), enabling finer-grained computation, and adding a low-cost interconnect for intra- and inter-mat data movement. Hardware modifications include isolated transistors, row decoder latches, and multiplexers to support MIMD execution and data reductions. The software stack provides compiler passes based on LLVM to automatically extract parallelism, schedule operations, and map data, along with OS support for address translation and coherence. Evaluation shows MIMDRAM achieves 37x better energy efficiency than CPU and 14x vs. prior PUD systems, and competitive throughput with 1% area overhead. The presenter highlights strengths, such as programmer transparency and leveraging LLVM, and weaknesses, including a simplistic cost model, poor handling of multiplication-heavy workloads, and limitations with floating-point arithmetic. The talk concludes with a Q&A session discussing optimization metrics and alternative compiler approaches.

185 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a comprehensive and balanced explanation of MIMDRAM, clearly outlining the problem, proposed solution, hardware/software co-design, and evaluation results. The argumentation is solid, grounded in the paper’s data and the presenter’s critical analysis. The speaker effectively demonstrates the value of finer-grained PIM and the importance of a holistic software-hardware approach to improve adoption. The discussion of strengths and weaknesses shows critical thinking, though the evaluation is largely based on the paper’s claims without independent verification.

Scientific Rigor, Source Quality, Title Accuracy

The presentation draws on a peer-reviewed paper (HPCA 2024) and includes references to related work, with links provided in the description. The technical content appears faithful to the original publication. However, the title’s mention of FC-DRAM is misleading as it is not discussed. The speaker acknowledges limitations and offers constructive critique, maintaining scientific rigor. No external validation was performed in the talk, but the breadth of sources (primarily the paper and recommended readings) supports reliability. The public comments (none provided) cannot be assessed; hence no analysis of audience feedback is included.

183 words

Title / Content Match

The title mentions both MIMDRAM and FC-DRAM, but the talk focuses exclusively on MIMDRAM. Some discrepancy exists, as FC-DRAM is not covered in the transcript.

Quality & Reliability

8/10

The presentation is based on a peer-reviewed HPCA 2024 paper, described accurately and critically. The speaker is knowledgeable, and the content aligns with the paper's findings. The talk includes personal analysis of strengths and weaknesses, maintaining scientific rigor.

Key Moments

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

Cited Sources

Concurring Sources

External References

Contribution & Novelties

The talk presents MIMDRAM as a novel hardware-software co-design that significantly improves the flexibility and programmability of processing-using-DRAM while maintaining high energy efficiency. The primary innovation is segmentation of DRAM wordlines into smaller mats, enabling fine-grained MIMD execution and low-cost inter-mat data movement. The software stack, built on LLVM, provides programmer transparency, which is a key step toward wider adoption of PIM. The presenter also contributes a critical analysis, identifying limitations in the cost model and handling of complex operations.

Pour aller plus loin :

  • Processing-in-memory — Background on the concept of integrating computation into memory.
  • SIMD — Contrast with MIMD execution models.
  • DRAM architecture — Overview of DRAM organization and operation.
  • LLVM — The compiler infrastructure used in MIMDRAM’s software stack.

122 words

Radar Profile

The radar profile shows high scores in technical level and information quality, reflecting the advanced topic and rigorous presentation. Quantitative information is also high, though slightly lower due to a focused scope on MIMDRAM. Reliability is strong based on the peer-reviewed source. Overall, the profile indicates a dense, technical seminar that accurately conveys research findings.

Reliability 8/10