Dear Md.Akibul Islam,

Resistive switching memory devices(ReRAM) have not become popular despite showing great promises over conventional RAMs because of following reasons :

01. Challenges in Analog Domain:

Working in analog domain brings several challenges, e.g., noise, non-zero wire resistance, nonlinear I-V characteristics, I/O stage resistance, etc. In addition, storing intermediary analog outcomes and implementing max pooling in analog domain is challenging. Further, on using analog circuitry, communication with the digital circuitry necessitates use of ADCs/DACs, however, these degrade signal precision and incur area/energy overheads. For example, ADC/DAC can take 85% to 98% of the total area/power of an neuromorphic computing system (NCS). Compared to this, digital signal transfer allows better control and high-frequency operation.

02. Reliability Challenges of ReRAM:

The high defect rate and process variation (PV) leads to reliability issues. For example, due to “single-bit failure”, a cell may get stuck at high or low conductance value, called stuck-at-one or stuck-at-zero (SA1/SA0), respectively. Especially for large NCSs, a ReRAM implementation leads to heavy wire congestion and poor reliability of read/write operations due to voltage-drop and PV. With increasing device failure rate, the accuracy of neural network (NN) reduces drastically. To mitigate this issue, redundancy-based techniques can be used, however, they incur complexity and area overheads.

03. Challenges in Achieving High Accuracy and Performance:

Compared to SRAM, ReRAM has high write energy/latency which increases the overall power consumption. ReRAM limitations, e.g., series line resistance and sneak-path, further reduce the performance. Further, during NN training, precise tuning of ReRAM requires frequent update of weights and large number of training iterations for convergence. This leads to high number of writes and large energy consumption. The non-ideal characteristics of ReRAM, e.g., PV and abrupt behavior during SET operation further increase the overhead of ReRAM tuning. Although the errors due to ReRAM faults or the analog-operation can be minimized by increased training, it leads to latency/energy penalty and aggravates ReRAM endurance issues. In addition, retraining may not be sufficient in case of high fault rate.

04. Limitations in Representing Neural Networks:

Not all Neural Network architectures/layers can be implemented using ReRAM, e.g., LRN (local response normalization) layers cannot be accelerated with crossbars.

I hope I have answered your question.

With Best Wishes,

Samir G. Pandya

Thank you very much Sir

RRAM is already being commercialized. Further improvements are expected for higher density. The relative immaturity, I.e, lack of clear model with observable effects, makes it hard for customers to take risk. At lower densities, however, products are already being produced to spec.

You are right. RRAM has already been commercialized. Despite its popularity in academia, major chip makers do not take it seriously because they do not regard it as a cash cow. Its potential market overlaps with the conventional memory market of the giants (DRAM and SSD). In the first place, the promise was the potential replacement of both DRAM and SSD as a storage class memory. However, now that the memory density of SSD is unbeatable, RRAM can't be a replacement of SSD by no means. Additionally, the state-of-the-art memory hierarchy with amazingly fast interfaces provides very fast virtual memory, so that RRAM is not much beneficial in terms of speed either.

I think control of variability is the main issue, all the parameters is not fulfill in a single material.

Doo Seok Jeong the 3D XPoint (an early type of RRAM) is already posing a challenge in SSD as a storage class memory candidate. Besides speed, the cell density is already superior to 3D NAND. 3D NAND is only competing on TLC and QLC, which compromises the reliability.

Muhammad Sultan Irshad the variability is addressed by ECC and redundancy as already done for NAND and DRAM.

ReRAM is already used in this product for example: https://www.adestotech.com/wp-content/uploads/RM24C128F.pdf

3D Xpoint is a PCRAM and was developed jointly by Intel and Micron. The products based on the Xpoint technology has been released in 2017 by Intel. These are Intel Optane memories.