PipeDRAM: A Data-Transposition-Free Processing-Using-DRAM Architecture with Hardware/Software Pipelining
PipeDRAM is a processing-using-DRAM (PUD) architecture that executes bulk bitwise and arithmetic operations directly on horizontally laid-out data. Prior PUD architectures (e.g., SIMDRAM) store operands in a vertical layout, with the bits of each element stacked along a DRAM column, and therefore have to transpose data between the horizontal layout used by the rest of the system and the vertical layout used for computation. PipeDRAM removes this transposition with two key ideas:
- Mat-aware bit mapping (MABM). A fixed bit permutation in the memory controller places all bits of a data element in one DRAM chip and maps its bit-position j to mat (j mod 32) of that chip. The permutation is applied to every cache line on writes and inverted on reads, so the CPU view of memory is unchanged.
- Pipelined execution. A modulo scheduler splits each PUD operation into bit-dependent primitives (on the carry chain) and bit-independent primitives, and overlaps them across mats, with each mat acting as one pipeline stage. Carries move between mats over a hierarchical in-DRAM copy path (LISA between subarrays, the inter-mat interconnect within a subarray).
This repository contains the simulation infrastructure, applications, and scripts used to evaluate PipeDRAM in our MICRO 2026 paper:
pud_simulator/: an analytical simulator that executes each PUD primitive (AAP/AP) of a uProgram individually. It models SIMDRAM, PipeDRAM (sequential, pipelined, and hierarchical carry propagation), Proteus, and a bit-parallel Kogge-Stone substrate, including data-transposition costs.workloads/: 11 real-world applications instrumented with an analytical PUD cost model.area/: Verilog for the MABM permutation and the area analysis.scripts/: one-command reproduction and validation of Figures 4 and 9-14.
./scripts/setup.sh
./scripts/reproduce.sh quick # all paper figures, under two minutes
./scripts/validate.sh quick- Getting started
- Reproducing the paper figures
- Repository structure
- Documentation
- Citation
- Acknowledgments
- License
- Ubuntu 24.04 (x86-64) with Python 3.12
- GCC/G++ and GNU Make (applications)
- Yosys and Icarus Verilog (MABM RTL checks)
- Poppler and
pdfcrop(figure post-processing and validation) - Liberation Sans (figure fonts)
sudo apt update
sudo apt install -y build-essential python3-venv texlive-extra-utils \
poppler-utils fonts-liberation yosys iverilog./scripts/setup.shThis creates .venv/ with the pinned Python packages from requirements.txt
and checks that the system tools above are installed.
source .venv/bin/activate
# List the 19 supported PUD operations
python -m pud_simulator --list-operations
# 32-bit addition on 64M elements in SIMDRAM, including worst-case transposition
python -m pud_simulator -op addition -n 32 -m 67108864 --transposition worst
# 32-bit multiplication in PipeDRAM with modulo scheduling and hierarchical
# carry propagation, using 32 subarrays
python -m pud_simulator -op multiplication -n 32 -m 67108864 \
--simulator pipedram --scheduling pipelined --carry-copy hierarchical \
--compute-subarrays 32
# Per-primitive trace of a 4-bit division
python -m pud_simulator -op division -n 4 --show-traceThe CLI uses the DRAM configuration of the paper (presets.pipedram_config()).
See pud_simulator/README.md for the Python API,
all options, and the DRAM parameters.
./scripts/reproduce.sh quick
./scripts/validate.sh quickquick mode regenerates Figures 4 and 9-14 and the area analysis from the
paper's application results in reference/results/ (under two minutes, less
than 2 GB of memory). full mode rebuilds and reruns all redistributable
applications at paper scale before plotting (up to 24 hours, 64 GB of memory,
10 GB of disk):
./scripts/reproduce.sh full [--heartwall-input /path/to/test_4k.avi]
./scripts/validate.sh fullFigures are written to generated/figures/, regenerated CSVs to
generated/results/, and logs to generated/logs/. A single figure can be
reproduced with ./scripts/reproduce_figure.sh <figure> replot|full, e.g.,
./scripts/reproduce_figure.sh 10 replot.
| Figure | Content | Generator | Output |
|---|---|---|---|
| 4 | Data-transposition overhead | plotting/plot_transposition_reuse.py |
transposition_reuse_combined.pdf |
| 9 | Bit-serial and bit-parallel throughput | plotting/plot_combined_scalability.py |
combined_scalability.pdf |
| 10 | Application execution time and energy | plotting/plot_speedup.py |
workload_normalized_combined.pdf |
| 11 | CPU-normalized energy efficiency | plotting/plot_energy_efficiency.py |
energy_efficiency_cpu_normalized.pdf |
| 12 | Mat and chip utilization | plotting/plot_mat_utilization.py |
mat_utilization.pdf |
| 13 | Bit-precision and data-size trade-off | plotting/plot_precision_datasize_tradeoff.py |
precision_datasize_tradeoff.pdf |
| 14 | Exposed carry-propagation overhead | plotting/plot_exposed_carry.py |
exposed_carry.pdf |
| - | DRAM area overhead | plotting/plot_area_comparison.py |
area_comparison.pdf |
PipeDRAM.ipynb walks through the same steps figure by figure.
Two inputs are not redistributed: the SPEC CPU2017 x264 source and media, and
the paper-scale heartwall video. Their paper results are included, so both
applications appear in all figures; full mode reruns heartwall when the
checksum-verified video is supplied. The kmeans input is generated
deterministically. See docs/DATASETS.md.
.
├── pud_simulator/ Per-primitive PUD simulator (SIMDRAM, PipeDRAM, Proteus, bit-parallel)
│ └── operations/ uProgram generators for the 19 PUD operations
├── workloads/ Instrumented applications (PolyBench, Rodinia, Phoenix)
│ └── util/ Analytical PUD cost model (bbop_manager.c/.h)
├── plotting/ Figure generators
├── area/ Area data and MABM Verilog (permutation, inverse, testbench)
├── scripts/ Setup, reproduction, and validation
├── tools/ Metrics collection, validation, area checks, cost-model table generator
├── reference/ Paper results, reference figures, and renderings
├── docs/ Reproduction, datasets, and area analysis
├── licenses/ Third-party notices
└── PipeDRAM.ipynb Figure-by-figure walkthrough
The cost-model lookup tables in workloads/util/bbop_manager.c are generated
from the simulator with python tools/generate_tables.py.
- Simulator: operations, options, Python API, DRAM parameters, and execution model.
- Workloads: instrumented applications and the PUD cost model.
- Reproducibility: what is reproduced, fixed inputs, and validation criteria.
- Datasets: kmeans generation and heartwall/x264 provenance.
- Area analysis: area values, their sources, and the MABM RTL checks.
If you use PipeDRAM, please cite:
@inproceedings{oliveira2026pipedram,
author = "Oliveira, Geraldo F and Olgun, Ataberk and Y{\"u}ksel, Ismail Emir and Bostanc{\i}, F. Nisa and Becker, Pedro H. E. and Sadrosadati, Mohammad and Ghose, Saugata and G{\'o}mez-Luna, Juan and Mutlu, Onur",
title = "{PipeDRAM: A Data-Transposition-Free Processing-Using-DRAM Architecture with Hardware/Software Pipelining}",
booktitle = "MICRO",
year = "2026"
}@misc{oliveira2026pipedram_arxiv,
author = "Oliveira, Geraldo F and Olgun, Ataberk and Y{\"u}ksel, Ismail Emir and Bostanc{\i}, F. Nisa and Becker, Pedro H. E. and Sadrosadati, Mohammad and Ghose, Saugata and G{\'o}mez-Luna, Juan and Mutlu, Onur",
title = "{PipeDRAM: A Data-Transposition-Free Processing-Using-DRAM Architecture with Hardware/Software Pipelining}",
howpublished = "arXiv:2609.30998 [cs.AR]",
year = "2026"
}The application instrumentation and the SIMDRAM/MIMDRAM cost models build on Proteus, and the uPrograms follow MIMDRAM. We thank the authors of PolyBench, Rodinia, and Phoenix, whose applications we instrument.
We acknowledge the generous gifts from our industrial partners, including Google, Huawei, Intel, and Microsoft. This work is supported in part by the ETH Future Computing Laboratory (EFCL), Semiconductor Research Corporation, AI Chip Center for Emerging Smart Systems (ACCESS), sponsored by InnoHK funding, Hong Kong SAR, and European Union's Horizon programme for research and innovation [101047160 - BioPIM].
PipeDRAM is released under the MIT License. The applications in
workloads/ keep their original licenses and notices (see
licenses/THIRD_PARTY_NOTICES.md).
Questions and bug reports are welcome as GitHub issues.