license: mit
library_name: pytorch
tags:
- verified-units
- neural-memory
- ddr5
- ecc
- retrocomputing
neural-ddr — verified DDR5 memory-logic units
Repositories: GitHub · 🤗 HuggingFace
Small neural nets that emulate DDR5 memory logic, each verified bit-identical to a golden reference over its entire finite input domain (N/N) — the same methodology as neural-aarch64-units.
A dev tool: present DDR5 behaviour (bus encoding, addressing, command decode, write-CRC, on-die ECC) on a machine that only has DDR3/DDR4. The bytes live in the host's real RAM; the neural nets emulate the logic.
Honest by design: no capacity or speed is invented. This models the data-path logic; it does not turn DDR3 silicon into DDR5.
Units
| unit | role | verification |
|---|---|---|
DBI |
data-bus-inversion encode/decode (DDR4/5) | 256/256 · 512/512 |
ADDR_MAP |
address → bank-group/bank/row/column | 4096/4096 |
CMD_DECODE |
control lines → DDR command | 32/32 |
WR_CRC |
write-CRC bit-slice (rippled) | 512/512 |
ODECC |
DDR5 on-die ECC — corrects a single-bit flip | 256/256 · 3328/3328 |
DDR5Bridge |
composes the units over real host RAM | 2048/2048 faults corrected |
Use
pip install torch
python step1_train_verify.py # DBI + memory bridge
python step2to6_train_verify.py # DDR5 units + ECC-correcting bridge
Weights: DBI.pt, DDR5_units.pt.
Create your own verified unit (template: neural_ddr/dbi.py): write the exact
golden finite function → enumerate the domain (decompose big/linear ones into
bit/byte slices, see crc.py/ecc.py) → common.train → common.verify must be
bit-exact on 100% of inputs → compose (ddr5_bridge.py).
Citation
@misc{byrne2026neuralddr,
title = {neural-ddr: Verified DDR5 Memory Logic Units},
author = {Byrne, Dean (Quazim0t0)},
year = {2026},
howpublished = {\url{https://huggingface.co/NeuralVerified/neural-ddr}}
}
Dean Byrne (Quazim0t0) · 2026
Now hosted by NeuralVerified.
This repo was moved into the NeuralVerified organization to help organize my profile. Originally published at
Quazim0t0/neural-ddr.