D, not sure I'll ever get used to ur new name!
Not sure if this has been posted, but would love ur take.
GLTA
Hi 66,
Well I'd never heard of this tech, so I'm not up to speed on the prospects.
ECRAM = ElectroChemical Random Access Memory.
As far as I understand it, the electrochemical bit refers to a what I would descrbe as the merging of an electrochemical capacitor (a bit similar to a battery) with a field effect transistor (FET). The transistor includes an electrolyte through which ions, eg, Li+, travel between source and drain under control of the gate voltage to change the charge state of the device.
https://onlinelibrary.wiley.com/doi/10.1002/smsc.202100006
Figure 7
However, it seems to be very much at a research stage:
https://en.wikipedia.org/wiki/Electrochemical_RAM
I would think it will not emerge as a practical device array for several years, if at all.
"
We will [therefore] be able to fabricate special purpose computer blocks (in say 5-10 years) where memory and transistors merge making them at least 1000 times more energy efficient than the best computers we have today for AI and simulation tasks (some calculation even show 1 million fold energy efficiency for certain algorithms).
We can likely expect the first commercial product to land before the end of this decade as the GTM (Go to Market) strategy requires at least five years of trials."
The comparison of 1000 times more energy efficiency than other ECRAMS only suggests that current ECRAMs are pretty slow because they say they need to scale to 2nm to become as fast as real transistors. Remembering that they are talking about a lab model, this means their current implementation is slower than real transistors.
Prof google shows that there is a lot of research in the ECRAM field.
As far as NNs are concerned, most of the talk is of ReRAMS.
https://onlinelibrary.wiley.com/doi/10.1002/smsc.202100006
4.3 Artificial Neurons and Artificial Synapses
CMOS integrated circuits used for simulation usually have a sundry circuit architecture, which is unlikely to be the ultimate solution for future artificial intelligence applications in terms of power consumption and size. Meanwhile, due to their conduction mechanism, which is similar to neural dynamics, and structure, which is similar to that of synapses, memristors are the most reported nonlinear storage devices used to simulate artificial synapses and artificial neurons.[
59]
Figure 11
The biological synaptic plasticity is an important parameter affecting the ability of synapses to achieve memory and learning. It is a measure of the synapse strength (corresponding to conductance of the device) as a function of the activity of neurons (electrical signals). Depending on the length of pulse duration, synaptic plasticity can be divided into short-term plasticity and long-term plasticity (STP and LTP). In the Ti3C2 artificial synapses, when a pair of adjacent spikes was applied, the postsynaptic current increased, and the depolarization-induced excitatory postsynaptic current increased sharply and then decreased in a short time, which is STP (Figure 11d). PPF is a polymerized form of STP. That is, in an artificial synapse, the application of one pulse will cause some ions to remain in the synapse, and then the second pulse will enhance the synaptic response. Similar to the behavior of biological synapses, artificial synapses can realize tunable synaptic weights by adjusting the pulse duration, number, amplitude, and frequency.
Figure 11e shows the spike-duration-dependent plasticity of the Ti3C2 artificial synapses.
Figure 11 cont ...
When the spike amplitude was fixed, the increase in the pulse time increased the number of ions that escaped from the surface of the MXene, resulting in a significant increase in the synaptic strength. By the same principle, when the number of spikes increased from 1 to 10, the synaptic weight of the Ti3C2 synapses increased sharply, exhibiting spike-number-dependent plasticity. However, a continuous increase in the number of pulses (≈50) caused the synapse weight to reach saturation and completed the STP-to-LTP conversion process. This is because as the number of spikes increased, a significant number of Li+ ions migrated to the Ti3C2-MXene/LPE surface, and some ions even escaped from the internal channel of Ti3C2. When the pulses were removed, the time for Li+ ions to return to their original position became longer, which means that the synaptic weight of Ti3C2 could be precisely adjusted by controlling the interval and sequence between the voltage pulses.
Escaping ions sounds like one problem they will need to fix.