Reducing Power Consumption of Embedded Dynamic Memories with ECCs

Andreas Burg; Wenqing Song; Yifei Shen

REVIEW 4 major objections 5 minor 16 references

Reducing Power Consumption of Embedded Dynamic Memories with ECCs

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that the minimum-power error-correcting code for retention-limited gain-cell DRAM is workload-dependent, shifting from strong codes in refresh-dominated regimes to light codes in access-dominated regimes, and reports 46.8–9

desk verdict Clean analytical study of ECC/refresh tradeoffs in GCRAM; the workload-dependent selection story is right, but the headline power numbers rest on an unvalidated refresh-power scaling law. read the letter →

arxiv 2607.16042 v1 pith:5SSECFGK submitted 2026-07-17 cs.IT cs.ARmath.IT

Wenqing Song , Yifei Shen , Andreas Burg This is my paper · ORCID

classification cs.ITcs.ARmath.IT

keywords GCRAMECCselectionrefreshintervalchipyieldpoweroptimizationBCHcodesretentiontimeembeddedDRAM

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that, for gain-cell embedded DRAM (GCRAM), the lowest-power memory design comes from jointly choosing the error-correcting code (ECC) and the refresh interval under a chip-yield constraint, rather than from picking the strongest code or the longest refresh interval. Using a log-normal retention-time distribution and a row-repair-aware yield model, the authors compute the longest feasible refresh interval for each BCH code, then a power model weighs refresh savings against parity-storage, access, and ECC-logic energy for a given workload. In a 16 MB case study they report that the optimal ECC shifts from strong codes when refresh dominates to light codes when access dominates, and that the selected configuration cuts total power by 46.8-94.8% relative to a no-ECC baseline. A careful reader would care because embedded memories increasingly dominate chip power, and this gives a quantitative rule for when ECC pays for itself.

What carries the argument

The key mechanism is the refresh-power scaling law (Eq. 11): refresh power is proportional to the physical bit count (including parity) times the ratio of the reference refresh interval to the chosen interval, tau0/tau_max. This law converts the yield-constrained refresh interval - computed from the log-normal data-retention-time model and the post-ECC row-failure probability - into a concrete power saving that can be traded against the extra parity storage and the synthesized BCH encoder/decoder energies. The row-repair-aware yield model (Eqs. 7-10) is what turns a seed retention distribution into a hard feasibility envelope per candidate.

What would settle it

Measure the actual refresh power of a GCRAM macro as a function of refresh interval on the same 16nm technology; if refresh power does not fall as tau0/tau_max (e.g., because of fixed per-refresh overhead or leakage), the model's refresh-power term is wrong. Alternatively, obtain the DRT parameters from the cited silicon measurement and reproduce the yield curves, then check whether the predicted refresh intervals (e.g., 165.87 us for ECC(552,512,4)) match a chip measurement.

Watch

Extended reading notes

Core claim

The central claim is that ECC strength and refresh interval should be selected together, not independently, for retention-limited GCRAM. The paper derives the refresh interval from a calibrated data-retention-time distribution plus a row-repair budget and a chip-yield target, and then selects the ECC candidate that minimizes total average power composed of refresh, memory-access, and encoder/decoder energy. The headline quantitative result is that, across bandwidth and activity sweeps, the selected ECC reduces total power by 46.8% to 94.8% compared with no ECC, with strong codes (BCH t=4) winning at low activity and light codes (BCH t=1) winning at high activity.

Load-bearing premise

The quantitative power savings rest on two calibrated inputs: the log-normal data-retention-time distribution (parameters not reported in the paper) and the assumption that refresh power scales inversely with the refresh interval (Eq. 11); if either is off - for example, if refresh power has a large fixed component or the retention tail is heavier than log-normal - the selected ECC and the 46.8-94.8% reductions change.

Editorial extensions

If this is right

The optimal ECC for an embedded GCRAM is workload-dependent: stronger codes win in low-activity, refresh-dominated regions, while lighter codes win in high-activity, access-dominated regions.
ECC can extend the feasible refresh interval by up to 21.3x (from 7.80 us no-ECC to 165.87 us with ECC(552,512,4)) at a 7.81% parity-storage overhead in the case study.
Total-power reductions of 46.8-94.8% are achievable across the workload sweep, with the largest reduction at the lowest bandwidth and activity.
Read/write composition matters: write-heavy workloads favor stronger codes because encoder energy is lower and more nearly constant than decoder energy across code strength.
Larger memory capacity favors stronger correction, because refresh power scales with capacity while ECC-logic energy is paid per access and does not scale directly with memory size.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The power model excludes leakage, arguing it is small and weakly ECC-dependent; if leakage becomes comparable to dynamic power at very low activity, the optimal ECC could shift toward weaker codes than the model predicts.
The inverse scaling of refresh power with refresh interval assumes all refresh energy is proportional to the number of refreshed rows; real controllers often have fixed per-refresh overheads, which would flatten the refresh-power curve and make the benefit of strong ECCs saturate.
The log-normal DRT parameters are taken from published silicon measurements but not reported in the paper, so an immediate test is to re-derive the yield curves with independently calibrated parameters and check whether the predicted refresh intervals (e.g., 165.87 us) reproduce on hardware.
The same yield-plus-power selection machinery should transfer to other retention-limited memories (conventional DRAM, FeFET, etc.) by recalibrating the retention model; applying it to published DRAM retention data would test whether the predicted ECC choices match deployed designs.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. The paper is a well-assembled analytical design study, not a chip measurement. It gives memory designers a workflow for choosing ECC strength in GCRAM under a given workload and yield target, and it argues convincingly that the right code moves from strong to weak as access activity grows. That qualitative claim is solid. The 46.8–94.8% power reductions should be treated as a model prediction with an upper-band flavor, because Eq. (11) assumes refresh power scales inversely with refresh interval and has no fixed overhead. If any constant cost exists in real refresh logic — controller clocking, precharge, leakage — the benefit of longer refresh intervals shrinks and the optimal ECC maps in Fig. 5 shift. The stress-test note is right: a 21× extrapolation from one reference point is taken without a measured P_refresh(τ) curve or circuit-level justification.

What is actually new is the activity-aware selection flow. The components (log-normal DRT, yield equations, BCH synthesis, average-power accounting) are known, but the combination and the demonstration that optimal ECC strength depends on workload is new and useful. The BCH encoder/decoder blocks are synthesized in 16nm and their energies come from gate-level simulation, which is real, reproducible evidence. Equations (1)–(15) are internally consistent; I traced the chain from pre-BER to yield to power and found no gap. Row-repair and yield targets are handled cleanly.

Soft spots beyond Eq. (11): the DRT parameters µ and σ are taken from [15] but not reported, so readers cannot reproduce the yield curves without hunting down that reference. There is no sensitivity analysis on µ, σ, P_ref0, or per-bit energies, which is a real omission for a decision-support tool. The yield model assumes independent retention failures across cells and rows; correlated process variation could change the numbers. Leakage is excluded, which the authors acknowledge, but they do not quantify how small it is. None of these kill the central message; they bound how literally to take the results.

Who it is for: memory subsystem designers working on embedded DRAM or retention-limited memories, especially in AI accelerators. It deserves a serious referee. The workload-dependent ECC selection story will almost certainly survive scrutiny; the quantitative claims need empirical grounding. I would recommend sending it to peer review and asking for a sensitivity analysis plus a measured or circuit-justified refresh-power curve before acceptance.

Referee Report

4 major / 5 minor

Summary. The paper proposes an activity-aware ECC-selection methodology for gain-cell embedded DRAM (GCRAM). It combines a log-normal data-retention-time (DRT) model, a row-repair-aware chip-yield model, and an average-power model that separates refresh power, memory access power, and synthesized BCH encoder/decoder power. The method selects a minimum-power ECC configuration under a chip-yield constraint. Using a 16 MB GCRAM case study with BCH codes from one- to four-bit correction, it reports that the optimal ECC shifts from stronger codes in refresh-dominated regimes to lower-overhead codes in access-dominated regimes and claims 46.8% to 94.8% total-power reduction versus a no-ECC reference. The equations in Sections III.A-III.C are internally consistent, and the optimization flow is clearly specified.

Significance. The paper's value is as a design-space exploration methodology rather than a new coding-theoretic or circuit-level contribution. Its strengths are the explicit closed-form power and yield models, the use of synthesized 16 nm BCH encoder/decoder energy data, and the clear workload-dependent selection maps. If the underlying refresh-power and DRT calibrations are accurate, the method is directly useful for memory-system designers. However, the headline quantitative claims rest on a small number of calibrated parameters and one load-bearing scaling assumption. The qualitative conclusion that stronger ECCs become attractive when refresh power dominates is a property of the model structure, so the contribution is primarily modeling, not an independent empirical finding. The quantitative reductions should be treated as predictions that require validation or at least sensitivity analysis.

major comments (4)

[§III-C, Eq. (11)] The inverse-linear refresh-power scaling is the main driver of the claimed 46.8-94.8% reductions. Eq. (11) assumes refresh power is purely dynamic and proportional to (L_phy/K)*(tau0/tau_max), with no fixed component, and it is extrapolated from the single calibration point tau0=100 us over a 7.8 us-to-165.9 us range (a 21x extrapolation). In the low-activity refresh-dominated regime of Fig. 6 (BW=512, alpha=0.02), the no-ECC reference is almost entirely refresh power, so even a modest fixed overhead (address generation, precharge, controller clocking, leakage) would substantially reduce the benefit of the longest refresh interval. Please provide measured or circuit-level P_refresh(tau) at several intervals, or add a fixed-power term to Eq. (11) and re-evaluate the selection and reduction percentages. Without this, the headline numbers are not yet supported.
[§IV-A, DRT calibration] The DRT parameters mu and sigma from Eq. (1) are not reported, and the calibration source [15] is cited only as a silicon measurement. The yield curves in Fig. 4, all tau_max values, and the entire ECC-selection map depend directly on p_tau. For reproducibility, report the calibrated mu and sigma (and the conditions under which they were obtained). In addition, provide a sensitivity analysis over reasonable ranges of these parameters (e.g., +/-10% in mu or sigma) to show that the qualitative selection regions and the claimed power-reduction range are stable. Without this, the quantitative conclusions are tied to unreported parameters.
[§III-C, leakage exclusion] Leakage is excluded from the power objective with the statement that it is 'relatively small' and weakly dependent on the ECC candidate. In a 16 nm technology, leakage is not negligible in a low-activity memory, and the regime where the paper reports the largest reduction (alpha=0.02) is precisely where leakage would be most visible relative to dynamic power. Moreover, stronger ECCs increase the physical row length L_phy, which increases array leakage. Please quantify leakage at the evaluated operating points or provide a circuit-level justification that leakage is small compared with P_refresh and Paccess. If leakage cannot be quantified, state that the reported absolute power numbers exclude it and show the sensitivity of the selection to a fixed leakage term.
[§III-B, Eqs. (5)-(7)] The yield model treats bit failures as independent Bernoulli events within a codeword and row failures as independent across rows and macros. Retention failures in gain-cell arrays are often spatially correlated due to process variation, which would alter the shape of the post-ECC yield curves and therefore the feasible refresh intervals. The paper provides no evidence or argument for the independence assumption. Please at least discuss the potential impact of correlated failures, or bound the error by comparing with a clustered-failure model. This is load-bearing because tau_max(c,u) and the subsequent power comparison inherit any error in the yield calculation.

minor comments (5)

[Abstract/Introduction] There are missing spaces in 'achieves46.8%to94.8%' and '46.8%to94.8%' in the abstract; also 'Section II provides' and similar minor spacing issues. These should be corrected.
[Fig. 3(e)] The legend says 'Darker bars: more injected errors i', but it is not immediately clear that for each t the bars correspond to i=0..t. Please state this explicitly in the caption or on the x-axis.
[Fig. 5 caption] The caption lists '1 ECC(136,128,1) ... 4 ECC(552,512,4)', but the figure panels use only numeric labels. Add the mapping directly to each panel or clarify the caption so the reader does not need to flip back and forth.
[§II-A, Eq. (1)] The log-normal DRT model is presented without a reference to the specific characterization data for Si-GCRAM. A citation to [14] is given, but the actual measured tail behavior should be briefly summarized, since the entire yield analysis depends on the tail of the distribution.
[§IV-B, yield crossing] The statement that all curves cross the 90% target at approximately 1.83e-3 post-ECC row-failure probability is useful, but it would be clearer to state that this value is exactly 1 - (0.9)^(1/(M*?)) under the independence assumption, or at least show the computation, because it helps the reader see the role of the repair budget.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ECC selection and power reductions are outputs of a stated model calibrated from external silicon data; the unvalidated refresh-scaling law is a validity concern, not a circular one.

full rationale

The paper's derivation chain is explicit and non-circular. The DRT model (Eq. 1) is a stated log-normal assumption calibrated from silicon measurements in [15]; the yield model (Eqs. 5-10) maps each ECC candidate to the largest refresh interval satisfying the yield target; the power model (Eqs. 11-15) combines refresh, access, and ECC-logic terms; Eq. 16 selects the argmin. The headline 46.8%-94.8% reductions and the Fig. 5 selection maps are computed outputs of these equations, not quantities that were fitted to produce the conclusion. No parameter is tuned to reproduce the selection maps or the reduction percentages; the only calibrated inputs are the DRT/log-normal parameters, the per-macro refresh reference power P_ref,0, and per-bit access energies, all stated as taken from [15]. The qualitative shift toward stronger ECCs at low activity is a consequence of the inverse-linear refresh-scaling assumption in Eq. (11) together with activity-proportional terms in Eqs. (12) and (14); it is a model implication, not a disguised reuse of an input. The self-citations ([13], [14], [15]) provide background and empirical calibration, not a uniqueness theorem or an ansatz smuggled in by authority. The main vulnerability of the paper is that Eq. (11) extrapolates the inverse-linear refresh-power law over a roughly 21x range (7.8 us to 165.9 us, Sec. IV-B) without presenting measured P_refresh(tau) data or accounting for constant refresh overheads; if that scaling is wrong, the optimal ECC and the reduction numbers would change. That is a correctness/validation risk, explicitly outside the circularity definition: the derivation is not equivalent to its inputs by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim depends on a chain of calibrated inputs: DRT distribution parameters, per-macro refresh power, per-bit access energies, and synthesized BCH energies. The paper reports some numeric values but leaves the DRT parameters in the cited reference, and no error bars are given. The modeling also assumes independence of bit/row failures and a 1/τ refresh-power scaling. These are the true prices the reader pays.

free parameters (4)

DRT log-normal parameters µ, σ = Not reported in paper; calibrated from [15]
The refresh-interval and yield results depend entirely on the retention-time distribution of Si-GCRAM cells. The paper states µ and σ can be calibrated from measured or published data and cites [15], but does not give values or error bars, so the quantitative power reductions are not reproducible from the paper alone.
Reference refresh power P_ref0 per macro = 112 µW at τ0 = 100 µs
Used in Eq. (11) to scale refresh power with interval and parity overhead; taken from [15], no error bar.
Per-bit read/write energies E_read, E_write = 20.8 fJ/bit, 12.8 fJ/bit
Used in Eq. (12) for access power; calibrated from [15], no error bar.
BCH encoder/decoder per-operation energies = Plotted in Fig. 3, values not tabulated
Synthesized at 16 nm; used in Eq. (14). Without tabulated values or code, an independent group must re-synthesize to reproduce the selection maps.

assumptions (6)

domain assumption GCRAM retention time T_ret is log-normally distributed (Eq. 1).
Section II-A. The log-normal model is taken from [14]; the paper does not justify it against measured distributions for the specific 16 nm GCRAM.
domain assumption Bit errors are independent across cells, codewords, rows, and macros (Eqs. 5-9).
Section III-B. This enables binomial/multinomial products; spatial correlation would alter yield and τ_max.
domain assumption Refresh power scales as P_ref0 * (L_phy/K) * (τ0/τ_max) (Eq. 11).
Section III-C. Assumes refresh energy per bit is constant and no fixed refresh overhead; a non-linear refresh-power law would change the optimum.
domain assumption Refresh bypasses ECC; errors are corrected only on system read.
Section III-A. This assignment determines that ECC logic energy is not charged to refresh; if refresh went through ECC or repaired errors, the power balance changes.
standard math BCH codes with capability t correct up to t errors in each codeword.
Section II-B. Standard coding-theory guarantee; the decoder is assumed to realize it.
domain assumption Leakage power is negligible and independent of the ECC candidate.
Section III-C. Ignoring leakage can bias the total power comparison, especially at low activity where refresh dominates.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reducing Power Consumption of Embedded Dynamic Memories with ECCs." pith.science (2026). https://pith.science/paper/5SSECFGK

@misc{pith2026260716042,
  author       = {Pith},
  title        = {Pith review of: Reducing Power Consumption of Embedded Dynamic Memories with ECCs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SSECFGK}},
  note         = {Machine review of arXiv:2607.16042}
}

read the original abstract

Gain-cell embedded dynamic random-access memory (GCRAM) offers dense and energy-efficient on-chip storage, but retention-time variations force frequent refresh operations to cover worst-case bits. Error-correction codes (ECCs) can alleviate this limitation by masking bit errors from weak cells and thereby reduce refresh cost. However, the trade-off between the additional access and logic energy introduced by ECCs and the power savings from longer refresh intervals is nontrivial, especially considering the wide range of available ECC options. To optimize overall power consumption, we propose an ECC selection method that combines a refresh-interval model with power analysis to identify the minimum-power ECC configurations under a given yield constraint. Across different memory bandwidths, activity factors, and read/write ratios, the evaluation results show that the best ECC option shifts from stronger codes in refresh-dominated operating regions to lower-overhead codes in access-dominated regions and achieves 46.8% to 94.8% reduction in total power relative to the no-ECC reference.

Figures

Figures reproduced from arXiv: 2607.16042 by the authors.

**Figure 3.** BCH implementation cost at 16 nm. IV. REFRESH–ECC TRADE-OFF EVALUATION A. Evaluation Setup and ECC Cost Characterization We evaluate the trade-offs based on the model in Sec. III for a 16 nm FinFET technology. For our case study, we consider a GCRAM memory with 16 MB of logical data, organized as M = 256 macros. Each macro has W = 512 rows, and each logical row stores K = 1024 data bits before ECC. Unless stated oth… view at source ↗

**Figure 4.** Post-ECC chip yield versus refresh interval of a [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: ECC selections under different workload conditions. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Total power across different workload points in the balanced case. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 2 linked inside Pith

[15]

A 512 kb 0.069µm 2 logic 3T GCRAM with 27µs retention time at 85 ◦c in 16 nm FinFET,

A. Yigitet al., “A 512 kb 0.069µm 2 logic 3T GCRAM with 27µs retention time at 85 ◦c in 16 nm FinFET,” inESSCIRC, 2025, pp. 293– 296

2025
[1]

A 667 MHz logic- compatible embedded DRAM featuring an asymmetric 2T gain cell for high speed on-die caches,

K. C. Chun, P. Jain, T.-H. Kim, and C. H. Kim, “A 667 MHz logic- compatible embedded DRAM featuring an asymmetric 2T gain cell for high speed on-die caches,”IEEE J. Solid-State Circuits, vol. 47, no. 2, pp. 547–559, Feb. 2012

2012
[2]

A 3T gain cell embedded DRAM utilizing preferential boosting for high density and low power on-die caches,

K. C. Chun, P. Jain, J. H. Lee, and C. H. Kim, “A 3T gain cell embedded DRAM utilizing preferential boosting for high density and low power on-die caches,”IEEE J. Solid-State Circuits, vol. 46, no. 6, pp. 1495– 1505, Jun. 2011

2011
[3]

GainSight: Application-guided profiling for composing heterogeneous on-chip memories in AI hardware accelerators,

P. Liet al., “GainSight: Application-guided profiling for composing heterogeneous on-chip memories in AI hardware accelerators,”arXiv preprint arXiv:2504.14866, 2025

arXiv 2025
[4]

OpenGCRAM: An open-source gain cell compiler enabling design-space exploration for AI workloads,

X. Wanget al., “OpenGCRAM: An open-source gain cell compiler enabling design-space exploration for AI workloads,”arXiv preprint arXiv:2507.10849, 2025

arXiv 2025
[5]

MCAIMem: A mixed SRAM and eDRAM cell for area and energy-efficient on-chip AI memory,

D.-T. Nguyen, A. Bhattacharjee, A. Moitra, and P. Panda, “MCAIMem: A mixed SRAM and eDRAM cell for area and energy-efficient on-chip AI memory,”IEEE Trans. VLSI Syst., vol. 32, no. 11, pp. 2023–2036, Nov. 2024

2023
[6]

A decoder for short BCH codes with high decoding efficiency and low power for emerging memories,

S. Choi, H. K. Ahn, B. K. Song, J. P. Kim, S. H. Kang, and S.-O. Jung, “A decoder for short BCH codes with high decoding efficiency and low power for emerging memories,”IEEE Trans. VLSI Syst., vol. 27, no. 2, pp. 387–397, Feb. 2019

2019
[7]

Fast energy optimization of on-chip ECC memories,

S. Rohman and F. Leduc-Primeau, “Fast energy optimization of on-chip ECC memories,” inSiPS, 2024, pp. 207–212

2024

Show all 16 references

[8]

Reducing cache power with low-cost, multi-bit error- correcting codes,

C. Wilkerson, A. R. Alameldeen, Z. Chishti, W. Wu, D. Somasekhar, and S.-L. Lu, “Reducing cache power with low-cost, multi-bit error- correcting codes,” inISCA, 2010, pp. 83–93

2010
[9]

SHIFT ECC: A value converting HBM ECC approach for refresh energy efficient integer quantized DNN inference,

J. Y . Lee, Y . S. Lee, Y .-H. Gong, S. W. Kim, and S. W. Chung, “SHIFT ECC: A value converting HBM ECC approach for refresh energy efficient integer quantized DNN inference,” inISLPED, 2025, pp. 1–7

2025
[10]

Stealth ECC: A data-width aware adaptive ECC scheme for DRAM error resilience,

Y . S. Lee, G. Koo, Y .-H. Gong, and S. W. Chung, “Stealth ECC: A data-width aware adaptive ECC scheme for DRAM error resilience,” in DATE, 2022, pp. 382–387

2022
[11]

CARE: Coordinated augmentation for elastic resilience on DRAM errors in data centers,

J. Chen, X. Jiang, Y . Zhang, L. Liu, H. Xu, and Q. Liu, “CARE: Coordinated augmentation for elastic resilience on DRAM errors in data centers,” inHPCA, 2021, pp. 533–544

2021
[12]

Modeling energy and delay for ECC-enabled cache architectures,

D. Joshi, A. Bagchi, A. Agarwal, G. Modi, N. Srivastava, S. Roy, and P. R. Panda, “Modeling energy and delay for ECC-enabled cache architectures,”IEEE Trans. Comput.-Aided Design Integr . Circuits Syst., Early access, 2026

2026
[13]

A 16-kB 65-nm GC-eDRAM macro with internal bias voltage generation providing over 100-µs retention time,

O. Harel, A. Yigit, E. Feifel, R. Giterman, A. Burg, and A. Teman, “A 16-kB 65-nm GC-eDRAM macro with internal bias voltage generation providing over 100-µs retention time,”IEEE J. Solid-State Circuits, vol. 60, no. 6, pp. 2239–2248, Jun. 2024

2024
[14]

Meinerzhagen, A

P. Meinerzhagen, A. Teman, R. Giterman, N. Edri, A. Burg, and A. Fish, Gain-cell Embedded DRAMs for Low-power VLSI Systems-on-chip. Springer, 2018

2018
[16]

Energy-efficient high-throughput VLSI architectures for product-like codes,

C. Fougstedt and P. Larsson-Edefors, “Energy-efficient high-throughput VLSI architectures for product-like codes,”J. Lightw. Technol., vol. 37, no. 2, pp. 477–485, Jan. 2019

2019

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [15]

A 512 kb 0.069µm 2 logic 3T GCRAM with 27µs retention time at 85 ◦c in 16 nm FinFET,

A. Yigitet al., “A 512 kb 0.069µm 2 logic 3T GCRAM with 27µs retention time at 85 ◦c in 16 nm FinFET,” inESSCIRC, 2025, pp. 293– 296

2025

[2] [1]

A 667 MHz logic- compatible embedded DRAM featuring an asymmetric 2T gain cell for high speed on-die caches,

K. C. Chun, P. Jain, T.-H. Kim, and C. H. Kim, “A 667 MHz logic- compatible embedded DRAM featuring an asymmetric 2T gain cell for high speed on-die caches,”IEEE J. Solid-State Circuits, vol. 47, no. 2, pp. 547–559, Feb. 2012

2012

[3] [2]

A 3T gain cell embedded DRAM utilizing preferential boosting for high density and low power on-die caches,

K. C. Chun, P. Jain, J. H. Lee, and C. H. Kim, “A 3T gain cell embedded DRAM utilizing preferential boosting for high density and low power on-die caches,”IEEE J. Solid-State Circuits, vol. 46, no. 6, pp. 1495– 1505, Jun. 2011

2011

[4] [3]

GainSight: Application-guided profiling for composing heterogeneous on-chip memories in AI hardware accelerators,

P. Liet al., “GainSight: Application-guided profiling for composing heterogeneous on-chip memories in AI hardware accelerators,”arXiv preprint arXiv:2504.14866, 2025

arXiv 2025

[5] [4]

OpenGCRAM: An open-source gain cell compiler enabling design-space exploration for AI workloads,

X. Wanget al., “OpenGCRAM: An open-source gain cell compiler enabling design-space exploration for AI workloads,”arXiv preprint arXiv:2507.10849, 2025

arXiv 2025

[6] [5]

MCAIMem: A mixed SRAM and eDRAM cell for area and energy-efficient on-chip AI memory,

D.-T. Nguyen, A. Bhattacharjee, A. Moitra, and P. Panda, “MCAIMem: A mixed SRAM and eDRAM cell for area and energy-efficient on-chip AI memory,”IEEE Trans. VLSI Syst., vol. 32, no. 11, pp. 2023–2036, Nov. 2024

2023

[7] [6]

A decoder for short BCH codes with high decoding efficiency and low power for emerging memories,

S. Choi, H. K. Ahn, B. K. Song, J. P. Kim, S. H. Kang, and S.-O. Jung, “A decoder for short BCH codes with high decoding efficiency and low power for emerging memories,”IEEE Trans. VLSI Syst., vol. 27, no. 2, pp. 387–397, Feb. 2019

2019

[8] [7]

Fast energy optimization of on-chip ECC memories,

S. Rohman and F. Leduc-Primeau, “Fast energy optimization of on-chip ECC memories,” inSiPS, 2024, pp. 207–212

2024