Skip to content
SASemiAgoraSemiconductor knowledge & engineeringOpen HBM partition

MEMORY REPLAY LAB / SA-MEM-002

Replay memory beyond SRAM.
Keep every label honest.

This lab collects the memory-expansion simulations for DRAM/HBM behavior, HBM/base-die partitioning, DRAM sensing, DRAM reliability guardbands, retention, refresh/ECC guardbands, RowHammer, NAND, 3D NAND, PCM drift, ReRAM variability, FeFET, MRAM switching, hierarchy placement, and CIM crossbar tradeoffs.

DRAM/HBM Bank1.606xMRAM Switch35.35 fJeNVM Margin866878 ppmCIM Crossbar0.122995DRAM 1T1C Sense126.565DRAM Refresh1072000RowHammer Risk112673.08NAND ECC Margin0.01FeFET Window441.6HBM Roofline0.3186Hierarchy Map0.76206PCM Drift223.588 mVReRAM Window164.612 mV3D NAND Vt/ECC16.82 dB

REPLAY CARDS

Different memories, different evidence.

Each card has a downloadable dataset, source spine, methods, limits, and public execution boundary. It is built for study and paper planning, not qualification, controller testing, or product ranking.

SA-SEED-MEM-DRAM-HBM-BANK-001

DRAM/HBM row-buffer and bank parallelism

Replay row-hit rate and bank count to see why locality and bank-level parallelism shape DRAM/HBM latency and throughput before raw bandwidth matters.

  • Compare row-hit rate before assuming more banks alone solve memory latency.
  • Read average latency and throughput index together so locality and parallelism stay coupled.
  • Use the result as a bridge to Ramulator2 or DRAMsim3, not as a DRAM-standard timing claim.
Open run card ->
SA-SEED-MEM-MRAM-SWITCH-001

MRAM switching energy versus retention

Replay STT/SOT-style write probability, pulse width, retention barrier, and write energy as a compact MRAM tradeoff card.

  • Move retention barrier and pulse width together to see why write energy is not independent.
  • Compare STT and SOT labels as architecture cues, not as calibrated stack predictions.
  • Use switch probability as a lesson proxy before studying micromagnetics or silicon data.
Open run card ->
SA-SEED-MEM-ENVM-MARGIN-001

RRAM, PCM, and FeFET read-margin degradation

Replay endurance, temperature, variability, and normalized read margin across RRAM, PCM, and FeFET style memory labels.

  • Treat emerging memory as a margin distribution problem rather than a single technology label.
  • Compare endurance and temperature before ranking RRAM, PCM, or FeFET.
  • Use the worst-case proxy to ask what metrology or compact model would replace this toy model.
Open run card ->
SA-SEED-MEM-CIM-CROSSBAR-001

CIM crossbar accuracy and energy tradeoff

Replay array size, ADC bits, conductance variation, accuracy proxy, and energy index for compute-in-memory crossbar intuition.

  • Compare array size, ADC precision, and conductance variation instead of treating CIM as a single knob.
  • Watch accuracy proxy and energy index move in opposite directions.
  • Use the card to prepare for CrossSim, NeuroSim, or MNSIM study without copying those tools.
Open run card ->
SA-MEM-003-DRAM-1T1C-SENSE-001

DRAM 1T1C charge sharing and sense margin

Replay cell capacitance, bitline capacitance, and stored voltage to see the small differential a DRAM sense amplifier must resolve.

  • Start from charge sharing before discussing DRAM timing names.
  • Compare cell and bitline capacitance to see why the initial signal is tiny.
  • Use sense margin and sense-time proxy as separate outputs.
Open run card ->
SA-MEM-003-DRAM-RETENTION-REFRESH-001

DRAM retention tail and refresh policy

Replay leakage class, temperature, and refresh interval to see why weak-cell tails drive refresh policy.

  • Look at typical, weak, and tail leakage classes instead of one average cell.
  • Trade refresh interval against bandwidth tax and fail-tail proxy.
  • Use temperature acceleration as a prompt for real retention characterization.
Open run card ->
SA-MEM-003-DRAM-ROWHAMMER-ECC-001

DRAM RowHammer disturb and ECC risk

Replay activation count, victim distance, and temperature to frame RowHammer as a pattern, margin, mitigation, and ECC-evidence problem.

  • Separate activation pattern, physical neighbor distance, temperature, and ECC residual risk.
  • Use victim margin as the cell-level bridge before system mitigation language.
  • Keep security, controller, and silicon evidence distinct.
Open run card ->
SA-MEM-003-NAND-THRESHOLD-ECC-001

NAND threshold distributions and ECC margin

Replay SLC/MLC/TLC/QLC threshold crowding, P/E cycles, retention time, and ECC margin proxy.

  • Compare bits per cell before treating Flash as one memory technology.
  • Read endurance and retention together because both widen threshold distributions.
  • Use ECC margin as a system-facing translation of device distribution overlap.
Open run card ->
SA-MEM-003-FEFET-HYSTERESIS-001

FeFET hysteresis, wake-up, fatigue, and usable window

Replay cycle count, temperature, and program pulse to see how ferroelectric memory window can wake up, fatigue, and shift.

  • Keep hysteresis window, imprint shift, wake-up, fatigue, and retention language together.
  • Compare early-cycle and high-cycle behavior before trusting one memory-window number.
  • Use the card to prepare for FeRAM/FeFET literature without copying device data.
Open run card ->
SA-MEM-003-HBM-THERMAL-ROOFLINE-001

HBM thermal bandwidth roofline

Replay stack height, target bandwidth, and cooling class to see usable bandwidth, energy per bit, and throttle risk together.

  • Do not read HBM bandwidth without energy per bit and thermal headroom.
  • Compare target bandwidth and usable bandwidth under throttle risk.
  • Use stack height as a package/system knob, not just a capacity slogan.
Open run card ->
SA-MEM-003-MEMORY-HIERARCHY-PLACEMENT-001

Memory hierarchy placement map

Replay SRAM, DRAM, HBM, SCM, NAND, and MRAM across workload weights to choose the right evidence question.

  • Pick the workload first, then compare memory technologies.
  • Separate latency, bandwidth, capacity, endurance, and energy movement instead of asking for one best memory.
  • Use the placement score as a discussion scaffold, not a product benchmark.
Open run card ->
SA-MEM-PCM-001

PCM SET/RESET pulse and resistance-drift replay

Replay SET/RESET pulse proxies, retention time, drift coefficient, and read margin so PCM is taught as a window-and-drift problem, not one ideal resistance ratio.

  • Read PCM margin through fresh and aged resistance windows, not just a SET/RESET label.
  • Move retention time and temperature together because drift turns a fresh window into a future read problem.
  • Compare margin and risk side by side before using PCM in a memory-hierarchy note.
Open run card ->
SA-MEM-RERAM-001

ReRAM set/reset variability and read-window replay

Replay compliance, cycling, temperature, selector on/off ratio, and LRS/HRS spread to see why ReRAM is a variability and array-read problem.

  • Treat HRS/LRS as distributions before treating ReRAM as a clean binary switch.
  • Keep selector leakage and cycle count visible because array read margin can collapse before the ideal device story does.
  • Use this as the deeper follow-up to the existing emerging-memory comparison card.
Open run card ->
SA-MEM-NAND3D-001

3D NAND string Vt distribution, read-retry, and ECC margin replay

Replay layer group, program/erase cycles, retention time, and read-retry steps to connect 3D NAND stack variation with threshold overlap and ECC margin.

  • Move from generic NAND threshold crowding into layer-aware 3D NAND behavior.
  • Keep retention, cycling, read-retry, and ECC in one table because storage reliability is a system chain.
  • Use layer group as a public-safe way to discuss stack nonuniformity without vendor data.
Open run card ->

Lab boundary

No live Ramulator2, DRAMsim3, NeuroSim, MNSIM, CrossSim, material compact-model, Flash ECC service, RowHammer attack workflow, account, upload, payment, comment, or signoff workflow is enabled. This is a static replay lab.

Open lesson packet ->