The on/off ratio of access devices: if this ratio is insufficient, then
the crossbar suffers from large power dissipation during programming and potentially also from write disturb. A reasonable
limit on crossbar size is the point when the total leakage in all
the unselected and half-selected devices matches the current
flowing through the selected device. For example, with an On/
Off ratio of $10^6$ , the limit on crossbar size is roughly $1000 x 1000$
Access device OFF-state leakage currents that are several orders of magnitude lower than the ON-state current allow the use of large arrays which are desirable to increase overall area efficiency. For instance, assuming that a single bit is being written in a $1000 x1000$ array, an access device ON/OFF current ratio of $10^6$ is needed in order to keep the
aggregate unselected cell leakage current to the same order of magnitude as the write current
ON/OFF ratio of MRAM is not very high, this will limit the size of the crossbar - did you estimate the limiting size?
This is actually a good point. No, we did not estimate it specifically.
If he meant a task of discerning a single ON against a background
of many OFFs, then indeed, 10000 OFFs might well be enough to
drown a few ONs in noise (or even 1000 OFFs would suffice). But
if we talk of general BNN performance based on the crossbar, we
saw it taking up to 10% of noise or so without much deterioration,
so the hopes there are much higher. But yes, we did not measure.
example - lets say $ R_{on} = 4 * R_{off} $ and say 5 ONs is among 10k OFFs, so the signal is $ 3* 5 $.
The relative signal is $ 15/10k $ which is 0.15 percent, and this relative signal should be higher than the level of electrical Johnson (or etc) noise.
## access device [page 12 of Xiao-Bennett](https://aip.scitation.org/doi/pdf/10.1063/1.5143815)
> **The on/off ratio of access devices:** if this ratio is insufficient, then
> the crossbar suffers from large power dissipation during programming and potentially also from write disturb. A reasonable
> limit on crossbar size is the point when the total leakage in all
> the unselected and half-selected devices matches the current
> flowing through the selected device. For example, with an On/
> Off ratio of $10^6$ , the limit on crossbar size is roughly $1000 x 1000$
2014 [ Burr et al.: Access devices for 3D crosspoint memory](https://researcher.watson.ibm.com/researcher/files/us-gwburr/JVSTB_2014.pdf) IBM
> Access device OFF-state leakage currents that are several orders of magnitude lower than the ON-state current allow the use of large arrays which are desirable to increase overall area efficiency. For instance, assuming that a single bit is being written in a $1000 x1000$ array, an access device ON/OFF current ratio of $10^6$ is needed in order to keep the
> > aggregate unselected cell leakage current to the same order of magnitude as the write current
## 2020 [A Variation Robust Inference Engine Based on STT-MRAM with Parallel Read-Out](https://knowen-production.s3.amazonaws.com/uploads/attachment/file/5177/luo2020.pdf) Yandong Luo
> Due to its limited on/off ratio, **STT-MRAM is often used as digital memory that only allows row-by-row read-out for near-memory computing.**
> We propose three techniques
> 1) a 2T-2MTJ bit-cell design with high on/off ratio,
> 2) redundancy for MSB weights to mitigate the impact of MTJ conductance variations, and
> 3) a hybrid-layer mapping scheme to reduce column current thus mitigating CSA offset effect
## 2016 [Resistive Memory Device Requirements for a Neural Algorithm Accelerator](https://www.osti.gov/servlets/purl/1365097) Sapan Agarwal
## ON/OFF ratio of MRAM is not very high, this will limit the size of the crossbar - did you estimate the limiting size?
- This is actually a good point. No, we did not estimate it specifically.
> If he meant a task of discerning a single ON against a background
> of many OFFs, then indeed, 10000 OFFs might well be enough to
> drown a few ONs in noise (or even 1000 OFFs would suffice). But
> if we talk of general BNN performance based on the crossbar, we
> saw it taking up to 10% of noise or so without much deterioration,
> so the hopes there are much higher. But yes, we did not measure.
example - lets say $ R_{on} = 4 * R_{off} $ and say 5 ONs is among 10k OFFs, so the signal is $ 3* 5 $.
The relative signal is $ 15/10k $ which is 0.15 percent, and this relative signal should be higher than the level of electrical Johnson (or etc) noise.