OpenOpt: An Open-Source SRAM Optimizer Based on Equivalent Circuit Model
Wang et al. · arXiv 2606.09129
The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.
Claim tested
Simulating an equivalent-cell circuit in place of the full SRAM array is many times faster (the paper's figure shows 8.1x at 64x64 rising to 61.4x at 256x512) with delay and power errors under about 2%.
What we ran and what came back
The released code at a pinned commit with the Xyce simulator on a 4-core cloud CPU, timing one read and write pass of the equivalent-cell circuit against the full array. The equivalent circuit was 6.2 times faster at 64x64 (bound 5.0; paper 8.1) and 12.4 times at 128x128 (bound 10.5; paper 17.9). At 256x128 it was 15.9 times faster with two replicates (interval 13.8 to 18.2), so the lower limit missed the bound of 15.0 and the verdict is Inconclusive. A follow-up decided after seeing that result, pre-registered separately and labelled post-hoc, gave 15.6 at 256x128 with six replicates (interval 15.0 to 16.3, narrowly above the bound) and 26.4 at 256x256 (bound 21.5; paper 35.9). It does not change the verdict. The 256x512 cell behind the paper's 61.4 times figure was not run, because the full array needs about 25 GB of memory. On accuracy, delay errors were at most 0.06% on the paper's table cells; one power figure (write static power at 32x16) was off by 5.6% at the 27 C we had chosen for that cell. A later rerun at 25 C, labelled post-hoc, gave -0.79%, which is the paper's figure, so the 5.6% reflects our choice of temperature and not the paper's cell; the same error was +3.6% at 26 C and +3.7% at 28 C. In the first run one negative control could not fire, which makes the accuracy arm Inconclusive.
What this result supports
In the pinned environment the equivalent circuit is faster than the full array on every cell run, by about 6 times at 64x64 and about 26 times at 256x256, and delay errors on the paper's table cells are below 0.1%.
What it does not support
The paper's speed-up figures (ours are about 63% to 76% of them), the 61.4x headline cell, which was not run, or the power-error bound on every cell or away from the reported temperature.
Record
Protocol, deposited before the run: 10.5281/zenodo.23552950, 10.5281/zenodo.23579224
The authors were sent this result. They have not reviewed this listing.
How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.