On the eve of the mass production of China's core, the new type of memory has entered the market in a big way. Is there a "replacement" effect?

Home > Sci-Tech

On the eve of the mass production of China's core, the new type of memory has entered the market in a big way. Is there a "replacement" effect?

2019-08-07 00:33:42 1553 ℃

The traditional memory technology has enabled the domestic Ziguang Group, Hefei Changxin, and Fujian Jinhua to compete in the competition. The domestic storage chip has replaced the import of the footsteps of the original fire, which has been difficult to calm down.

However, another new memory technology team that has been squatting for nearly 20 years, including MRAM, PCRAM and ReRAM, has benefited from key breakthroughs in technology, materials and equipment, and is on the way to mass production. We are at a turning point in witnessing the history of memory.

However, at this point in time, it is also the impact of the new memory technology on traditional memory DRAM, 3D NAND, SRAM. Will it form a "replacement" effect?

Intel 3D XPoint turned out, industry rekindling hope

New memory can be divided into stand-alone products, and embedded in logic technology to replace some of the traditional embedded flash memory eFlash technology And in embedded technology, the trend has matured rapidly. However, for stand-alone memory, there are still performance and cost issues to be overcome.

Therefore, the new memory, whether it is MRAM, PCRAM and ReRAM, will not impact the DRAM and 3D NAND chip industry, which is now in full swing in China, but it is driven by some application fields such as cloud computing and Internet of Things. Edge computing, with the addition of new memory technology, can really make the development of the entire industry even more powerful.

Graph | 3D XPoint (Source: Intel)

New memory technology has been proposed for nearly 20 years, mature The road is stumbling. Until 2015, Intel's 3D XPoint technology was born, and it is considered to be similar to the structure of PCRAM. The whole new storage technology is suddenly clear, and the development in the next few years is even more powerful.

In order to add firewood to the new memory industry, the application materials for global semiconductor leaders are for MRAM, PCRAM, ReRAM introduced two machine equipment: Endura Clover MRAM physical vapor deposition (PVD) machine, and Endura Impulse physical vapor deposition (PVD) machine, which has become a powerful driving force for the development of the industry.

DeepTech and the two experts in application materials, Dr. Zhao Ganming, General Manager and Chief Technology Officer of Applied Materials China, and Dr. Zhou Chunming, Global Product Manager of Applied Materials Metal Deposition Products, gave a glimpse What kind of changes the new memory will bring to the world will witness the historical turn of the storage industry.

Here, DeepTech's panoramic analysis of the key reasons for the recent rise of new types of memory, which manufacturers have begun mass production, application principles and areas, and the benefits to the industry.

Moore's Law is fading, new memory is on the battlefield

Moore's Law, which was introduced in 1965, has been written for more than 50 years, writing countless milestones for the global electronics industry, but today, The chips designed and manufactured according to this law are gradually declining in the four major standards of PPAC (power consumption, performance performance, area Area, cost Cost).

Many IoT and cloud computing chips are not available from Moore's Law. Why?

In the context of the “Internet of Everything” and “Industry 4.0” eras, data has exploded. For example, we generate about 1GB of data per day, but when you drive a driverless car, the amount of data generated in a day can be as high as 4000 GB, which is equivalent to 4,000 times.

2019 is a crucial year. The data generated by the machine has exceeded the data generated by humans. This is the first time in human history; it is expected that by 2022, the data generated by the machine may be human. Produce as much as 9 times the data.

(Source: Pixabay)

The logic of future world computing is that data comes from the collection of machines, including cars. ,Smart City,For smart homes, all generated data must be transmitted and calculated from the terminal and from the edge through the layers, then to the cloud, to the big data center, and then back to the terminal.

In this short period of time, after the influx of data, the source of data is constantly being calculated, processed, and retransmitted, which is very challenging for chip performance, and the existing computing architecture has long been Unable to meet core needs.

In the past "Moore's Law" era, the pursuit of transistors is getting smaller and smaller, the goal is to double the number of transistors every 18 months to two years, but with the effect of the law decreasing, from 14 From nm nanometers to 10 nm nanometers, it may take 4 years, from 10 nm down to 7 nm, 5 nm, it takes longer. Imagine that Intel 10 nm has been deferred to get a glimpse of it.

Therefore, more and more people are arguing whether Moore's Law has reached the end of its life?

To answer this question, we can think like this. If you continue to use traditional thinking and increase the density by reducing the size of the transistor, you can't do it. However, from another perspective, there are many ways to achieve the above PPAC (power consumption, performance, area, cost).


The industry has now proposed various "tricks" to continue Moore's Law, This DeepTech is explored on five levels.

New architecture: Google's TPU, Nvidia's GPU, is an accelerator role to improve computing, especially in the cloud.

New structure: such as 2D two-dimensional to 3D three-dimensional NAND, because it is a three-dimensional structure, so you can always go up, the space up can be improved.

New material: I used to take out a periodic table of elements. The elements in semiconductors and transistors have been added a lot. For example, the copper process replaces the aluminum process, and the cobalt replaces the copper. Improve the performance of the transistor and highlight the important role that new materials play in improving PPAC.

New Miniature Technology: ASML Extreme UV Lithography EUV.

New packaging technology: Different kinds of different process technologies, such as 28nm or 5nm processors, memories, accelerators and other chips, are integrated through advanced packaging technology to achieve optimal system level. Performance. For example, the packaging technologies InFO, CoWoS, and 3D ICs introduced by TSMC in recent years, as well as EMIB and Forevos launched by Intel, are all started from the latter stage and continue the life of Moore's Law.

In-Memory Computating concept fires up

The era of big data highlights the importance of huge computing needs, but also drives the development of hardware and the revival of investment, such as the accelerator mentioned above, if To go deeper into the discussion, here is a concept, "Near Memory Computing."

What is Near Memory Computing? To put it simply, in the past we often had the idea that "the processor is king" and that the power of the processor is the most important, but not now.

Because computing power is no longer a separate processor capability decision, but rather the data is transferred back and forth between the processor and the memory, and because of the bottleneck, the computing power can no longer advance.

The definition of Near Memory Computing is to use a large amount of high bandwidth and large capacity to more closely connect the memory and the computing processor to increase the computing performance at the system level.

This concept is actually implemented with existing building blocks, such as DRAM, NAND, SRAM, etc., and will gradually be combined with new memory MRAM, ReRAM, PCRAM to increase computing performance and create "memory computing". "The basis of (In-Memory Computating).

In-Memory Computating has been a very hot concept in recent years, but it may take at least 3-5 years to achieve. Different from Near Memory CompuTing is to put storage and processing closer together. In-Memory Computating integrates storage and processor for calculation, there is no transmission, delay, etc., and the performance is greatly improved.

Looking back 10 to 20 years, brain-like calculations and quantum calculations can achieve the above goals, but these technologies are too far-reaching. If you want to achieve the goal of In-Memory Computating as soon as possible, at least within 5 years. New memory will play a very important role.

Which semiconductor manufacturers have begun mass production of new memories

Before discussing the operation of new memories, let's talk about which semiconductor manufacturers have mass-produced MRAM, ReRAM and other technologies.

The current camps that are investing in R&D or production of new memory technologies can be divided into three broad categories.

The first category: logic process foundries, including TSMC, GlobalFoundries, SMIC, Samsung Electronics, etc., mainly embedded in the mainstream process technology MRAM, ReRAM storage technology, belonging to embedded memory Used, not a stand-alone memory.

Category 2: Stand-alone memory manufacturers, such as Group and Evenspin, have integrated 1Gb STT-MRAM into an enterprise-class SSD system to act as a cache to improve SSD performance.

The third category: research institutions, academic units, etc.

In addition to the 3D XPoint technology developed by Intel and Micron, the semiconductor companies that are developing MRAM, ReRAM, and PCRAM technologies include TSMC, IBM, SK Hynix, Western Digital, and GlobalFoundries.


TSMC has disclosed the MRAM and ReRAM technology processes in the technical forum.

TSMC's current 40nm ReRAM has mass production capability in the IoTOn the network chip, it replaces the traditional embedded flash eFlash technology, emphasizing that the stored chip can be stored for 10 years and after 10,000 times of reading and writing.

Furthermore, TSMC's 22 nm MRAM also has mass production capabilities. Unlike ReRAM technology, this MRAM technology is used in mobile devices, high-performance computing HPC, automotive electronics, etc. Embedded Flash eFlash technology.

In terms of performance, 22 nm MRAM is three times faster than eFlash technology, and the data can be stored for 10 years and subjected to 1 million readings and readings at high temperatures.

Sung Gon Jin, head of SK Hynix Advanced Thin Film Technology, also said that in addition to DRAM and NAND, the development of next-generation memory is also being introduced to improve the efficiency of the data center.

In addition, GlobalFoundries is also investing in new state of the art embedded memory technology MRAM for many years, jointly developed with Everspin, the company has also revealed the introduction of embedded eMRAM technology in the 22 nm FD-SOI process to produce complex cars. With the MCU chip, it will be used in advanced driver assistance systems (ADAS) systems or in other automotive systems.

Graph | GlobalFoundries (Source: DeepTech)

Principles and Applications of New Memory

MRAM is a magnetic random access memory. The architecture is that the memory cells in the transistors are interconnected at the back end, even without occupying the "silicon" area, and can be directly embedded in the logic circuit, so it can be done very small. One transistor and one memory cell.

Furthermore, PCRAM is a phase change random access memory, and ReRAM is called a resistive random access memory. What is more attractive than MRAM is that these two new storage technologies can be implemented like NAND. 3D 3D architecture.

The advantage of the 3D architecture is that it can be stacked all the time. When adding one layer, the density of the memory can be doubled. In addition, the cost can be reduced. Such characteristics can be used for large capacity and low cost. Cloud computing and big data centers are very attractive.

It can be said that the new memory has a wide range of applications, but if the benefits are maximized, the two applications will be locked: Internet of Things, cloud computing and big data centers.

The Internet of Things we often talk about is the so-called edge terminal and edge device.

The current edge device architecture is a logic chip plus an SRAM chip. The function of the SRAM is to calculate and then add a 3D NAND chip to store the algorithm/software/code.

The so-called "edge" is because there is no connection and no power. At this time, the power consumption problem is very important, because the power consumption can determine how long it takes.

At this point, MRAM can replace the function of SRAM. Because SRAM is also power-consuming and even leaking when not in use, some edge devices may be on standby for 99% of the time. If you replace the SRAM with MRAM, you can improve many power consumption problems.

The same is true for 3D NAND. It is actually a high-voltage device. If part of the MRAM is used instead of 3D NAND, the power consumption can be reduced.

MRAM has two major advantages. The first is that it does not consume power when it is in standby. The second is much cheaper than flash memory. If it is shortcoming, the speed of MRAM has not yet reached the SRAM level. For example, MCUs that are widely used in the Internet of Things, MRAM is very suitable for use.

Next, look at the cloud and the big data center. There are three challenges in this area. First of all, the influx of massive data, and then the need to perform fast calculations, the third key is still to return to power consumption.

The current mainstream architecture is DRAM plus SSD to store data, but how do you use new memory to improve performance?

Method one is to replace the DRAM part, because from the power point of view, DRAM has power consumption toproblem. Furthermore, PCRAM and ReRAM can be used as a 3D architecture and have an advantage in cost.

Method 2 is to replace the SSD part. The advantage of SSD is that it is cheaper. Thanks to the maturity of 3D NAND stacking technology, 128-layer stacks are now in mass production. The cost of 3D NAND is getting lower and lower, but the weakness is performance.

If you replace some DRAM with PCRAM and ReRAM, you can also implement 3D architecture, and performance is much better than SSD.

(Source:Applied Materials )

How does the new memory work

Magnetic memory is a The three-layer structure, called the "tunnel junction" in the middle, is magnesia, two magnetic layers on both sides, the magnetic layer can be understood as two magnets, and the two magnets have north and south poles. If the north and south poles match, the electrons It is easy to pass, and the resistance is a relatively low state.

Furthermore, the magnetic layers on the upper and lower sides can be turned parallel to the lower side by the current, which is a mismatch. When there is no match, the electron is difficult to pass, and it is a high resistance state.

Therefore, the realization of "0" and "1" storage through low resistance and high resistance is actually a memory technology based on resistance change, which realizes high resistance and low resistance by magnetism.

PCRAM, ReRAM are similar in principle and are controlled by current or voltage. PCRAM is a crystal form to control low resistance and high resistance. When it is fully crystallized, it is a low-resistance state. When it is amorphous, it is a state of high resistance, thereby realizing "0" and "1".

ReRAM is similar. The non-conducting place is a high-resistance, just like the insulating material. After the power is turned on, the conductive path can be realized, showing a low-resistance state, so similar to the MRAM, through the high and low resistance To achieve "0" and "1".

In short, to achieve this new type of memory is to achieve the basis of these memories through material engineering, there are still some challenges to overcome.

(Source:Applied Materials)

Equipment technology breakthrough, the era of scale has finally arrived

For large-scale production of new types of memory, material engineering breakthroughs in equipment manufacturers are key. Application Materials For the MRAM-designed Endura Clover MRAM PVD system, multiple process steps can be performed under vacuum, implementing 10 materials of the entire MRAM, and then stacking 30 layers by layer, the core of which is Clover PVD, one The chamber can be used to achieve up to five materials and then deposit a thin film at the atomic and subatomic levels.

It is mentioned that there is a magnesium oxide layer in the middle of MRAM. The application material indicates that the middle magnesium oxide layer is very important, which will affect the performance of the entire MRAM Device. The application materials are built using unique techniques to make the entire MRAM realize. Low power consumption and high durability.

In the MRAM manufacturing process, it is very complicated to realize the deposition and deposition of more than 10 materials on more than 30 layers on one platform. In contrast, PCRAM and ReRAM are not so many layers, but it still has many layers of structure, including electrodes, selectors, and memory. The materials inside are very unique.

For example, PCRAM, whose material structure is GST, contains 锗 Ge, 锑 Sb, 碲 Te, is not a commonly used material, the challenge is how to deposit these composite materials and control their composition.

For mass production of PCRAM and ReRAM, the material corresponding to the application material is Endura Impulse PVD system, which can strictly control the composition of multi-component materials, and achieve excellent film thickness, uniformity and interface control.

In the general trend, large-scale mass production of new memories will start from the embedded stage. For example, TSMC embeds ReRAM and MRAM into existing processes, and then new storage technologies will develop into independent memory fields because It needs more density.

In the era of "data explosion", the chip urgently needs high computing performance, but it is in the era of slowing down Moore's Law, and the brain-like chip and quantum computing distance are too far. The new storage technology has been grinding for many years. In the event of a breakthrough in equipment materials, it is possible to catch up with the era of interconnected objects and massive data calculations, and fight on the battlefield.

On the occasion of large-scale mass production of new-type memory, it is just when the domestic traditional 3D NAND and DRAM memory enter the international competition arena. Although the fields of application and level are different, but coincidentally, the old and new technologies are also on the page of historical turning point, witnessing the trajectory of the global technology industry.