075582814553
In our daily life, there are actually far more than three storage media: HDD hard disk, floppy disk and CD-ROM. The U disk, TF card, SD card that we often use, as well as the DDR memory and SSD hard disk used in the computer are all other storage technologies. This technology, we call it "semiconductor storage". Today, I will focus on telling you about this knowledge.
Modern storage technology, in general, is divided into three major parts, namely magnetic storage, optical storage and semiconductor storage.
Figure. 1
Semiconductor memory, in short, is a memory that uses a "semiconductor integrated circuit" as a storage medium.
If you disassemble your U disk or SSD hard disk, you will find that there are PCB circuit boards, as well as various chips and components. Among them is a type of chip, which is dedicated to storing data, and is sometimes called "memory chip".
Figure. 2 The Structure of SSD Hard Disk
Compared with traditional magnetic disks (such as HDD hard disks), semiconductor memory is lighter in weight, smaller in size, and faster in reading and writing. Of course, the price is also more expensive.
In recent years, the whole society has paid a lot of attention to the chip semiconductor industry. However, what everyone mainly focuses on is computing chips such as CPU, GPU, and mobile phone SoC.
As everyone knows, semiconductor memory is also one of the core pillars of the entire semiconductor industry. In 2021, the global semiconductor memory market will be worth $153.8 billion, accounting for 33%, or one-third, of the entire integrated circuit market.
Figure. 3 The proportion of major semiconductor categories in the world in 2022 Memory is down, but still 26%
Semiconductor memory is also a large category, which can be further divided into: volatile (VM) memory and non-volatile (NVM) memory
Figure. 4
As the name suggests, volatile memory cannot retain data when the circuit is powered off, while non-volatile memory can.
This is actually easier to understand. Children's shoes who have learned the basics of computers should still remember that storage is divided into memory and external memory.
Memory used to be called running memory (running memory). After the computer is powered on, it works with the CPU. After the power is turned off, the data is gone and belongs to volatile (VM) memory.
The external storage, that is, the hard disk, stores a large number of data files. When the computer is shut down, as long as you perform a save (write) operation, the data will continue to exist, belonging to non-volatile (NVM) memory.
Please pay attention: Many materials now also divide semiconductor memory into random access memory (RAM) and read only memory (ROM). You should be familiar with it, right?
Figure. 5
ROM read-only memory: well understood, can be read, not written.
RAM random access memory: It means that it can "randomly read or write data from any storage unit of the memory", which is relative to traditional magnetic storage, which must be "sequential access (Sequential Access)".
Some people think that volatile memory is RAM and non-volatile memory is ROM. In fact, this is not rigorous, and the reason will be discussed later.
In the past few decades, volatile memory has not changed significantly, mainly divided into DRAM (Dynamic Random Access Memory, Dynamic RAM) and SRAM (Static Random Access Memory, Static RAM).
Figure. 6
DRAM
DRAM is composed of many repeated bit cells (Bit Cell), each basic cell is composed of a capacitor and a transistor (also known as 1T1C structure). The amount of charge stored in a capacitor, used to represent "0" and "1". The transistor is used to control the charging and discharging of the capacitor.
Figure. 7
There will be leakage due to capacitors. Therefore, periodic "dynamic" charging must be performed to maintain the potential before data changes or power outages. Otherwise, data will be lost.
Therefore, DRAM is called "dynamic" random access memory.
DRAM has always been the mainstream solution for computer and mobile phone memory. Computer memory (DDR), graphics card memory (GDDR), and mobile phone operating memory (LPDDR) are all types of DRAM. (DDR basically refers to DDR SDRAM, double-rate synchronous dynamic random access memory.
It is worth mentioning that, in addition to GDDR, there is a new type of video memory called HBM (High Bandwidth Memory). It is formed by stacking many DDR chips and encapsulating them together with the GPU (no memory particles can be seen on the outside).
SRAM
SRAM may be unfamiliar to everyone. In fact, it is the technology used by our CPU cache.
The structure of SRAM is much more complicated than that of DRAM.
The basic unit of SRAM consists of at least 6 transistors: 4 FETs (M1, M2, M3, M4) form two cross-coupled inverters, and 2 FETs (M5, M6) are used for reading The control switch of the written bit line (Bit Line) forms a latch (flip-flop) through these field effect transistors, and locks the binary numbers 0 and 1 when powered on.
Therefore, SRAM is called "Static Random Access Memory".
Figure. 8
SRAM memory cell
SRAM does not need to be refreshed regularly and has a fast response speed, but it consumes a lot of power, has low integration and is expensive.
So，it is mainly used for the main cache of the CPU as well as the auxiliary cache. In addition, it will also be used in FPGA. Its market share has always been relatively low, and its presence is relatively weak.
Next, look at non-volatile memory products.
There are many technical routes for non-volatile memory products. The earliest is the ROM mentioned earlier.
The oldest ROM, that is the "real" ROM - completely read-only, when it leaves the factory, the contents of the storage are already hard-coded and cannot be modified in any way.
This kind of ROM has poor flexibility. If there is something wrong, there is no way to correct it, and it can only be discarded.
Mask read-only memory (MASK ROM) is the representative of the above ROM. To put it bluntly, it is to directly use the mask process to "engrave" the information into the memory, so that the user cannot change it, which is suitable for early mass production.
Later, experts invented PROM (Programmable ROM, programmable ROM). Such ROMs can generally only be programmed once. When shipped from the factory, all memory cells are set to 1. Through special equipment, the fuse can be blown by current or light (ultraviolet) to achieve the effect of rewriting data.
The flexibility of PROM is higher than that of ROM, but it is still not enough. It is best to be able to modify the data, so some experts invented EPROM (Erasable Programmable, erasable programmable ROM).
The method of erasing can be light or electricity. Electricity is more convenient. Erasing with electricity is called EEPROM (electrically erasable programmable EEPROM).
Figure. 9
EEPROM takes Byte as the smallest modification unit. That is to say, you can write 0 or 1 to each bit, that is, read and write according to the "bit", and you do not need to erase all the content before writing. Its erase operation is also based on "bit", and the speed is still too slow.
In the 1980s, Toshiba's technical expert, Fujio Okaoka, invented a new memory that can quickly erase operations, that is, Flash (flash memory).
Flash means "quickly" in English.
Due to space limitations, we will introduce the specific principles of FLASH next time. We just need to know that Flash storage is erased in "blocks".
Common block sizes are 128KB and 256KB. 1KB is 1024 bits, which is several orders of magnitude faster than EEPROM erasing by bit.
At present, there are only two mainstream representative products of FLASH, namely: NOR Flash and NAND Flash.
Figure. 10
NOR Flash
NOR Flash is a code-type flash memory chip, and its main feature is the execution in the chip (XIP, Execute In Place), that is, the application does not need to read the code into the system RAM, but can run directly in the Flash flash memory.
Therefore, NOR Flash is suitable for storing codes and some data, with high reliability and fast reading speed, and has advantages in performance and cost in low- and medium-capacity applications.
However, the writing and erasing speed of NOR Flash is very slow, and its volume is twice that of NAND Flash, so its use is limited and its market share is relatively low.
In the early days, NOR Flash was also used in high-end mobile phones, but later, after the introduction of eMMC in smartphones, even this market was crowded out.
In recent years, the application of NOR Flash has rebounded and the market has recovered. NOR Flash is widely used in low-power Bluetooth modules, TWS headsets, mobile phone touch and fingerprints, wearable devices, automotive electronics, and industrial control.
NAND Flash
In contrast, the market share of NAND Flash is much larger.
NAND Flash is a data-type flash memory chip that can realize large-capacity storage.
It reads and writes data in units of pages and erases data in units of blocks. Therefore, although its writing and erasing speed is about 3-4 orders of magnitude slower than DRAM, it is also 3 orders of magnitude faster than traditional mechanical hard disks. For eMMC/EMCP, U disk, SSD and other markets.
eMMC was mentioned earlier. In the past few years, the term was still quite popular.
Figure. 11 eMMC
eMMC is an embedded Multi Media Card. It encapsulates the MMC (Multimedia Card) interface, NAND and the main controller in a small BGA chip, mainly to solve problems such as NAND brand differences and compatibility. Manufacturers introduce new products quickly and easily.
The eMCP is a package of eMMC and LPDDR, which further reduces the module size and simplifies the circuit connection design.
In 2011, the UFS (Universal Flash Storage, Universal Flash Storage) 1.0 standard was born. Later, UFS gradually replaced eMMC and became the mainstream storage solution for smartphones. Of course, UFS is also based on NAND FLASH.
Figure. 12 The Standard Configuration of Mainstream Mobile Phones in Recent Years
SSD, everyone should be familiar with it. It basically uses NAND chips, and is currently developing very rapidly.
Figure. 13 Internal structure of SSD
According to the difference in internal electronic unit density, NAND can be divided into SLC (single-layer memory cell), MLC (double-layer memory cell), TLC (three-layer memory cell, QLC (four-layer memory cell), representing each memory cell in turn The stored data are 1-bit, 2-bit, 3-bit, and 4-bit respectively.
From SLC to QLC, the storage density will gradually increase, and the cost per bit will also decrease. But relatively, performance, power consumption, reliability and P/E cycle (the number of erase and write cycles, that is, life) will decrease.
In recent years, there has been a lot of controversy surrounding SLC/MLC/TLC/QLC in the DIY installation circle. At first, netizens felt that the lifespan of SSD hard drives would shrink. Later, it was found that the shrinkage was not so serious, and the lifespan was still sufficient. So, it was slowly accepted.
The early NAND were all 2D NAND. After the process enters 16nm, the cost of 2D NAND rises sharply, and the difficulty and cost of the planar scaling process are unbearable. Thus, 3D NAND appeared.
Figure. 14
To put it simply, from the bungalow to the building, the three-dimensional stacking is used to increase the memory capacity and reduce the process pressure of 2D NAND.
In 2012, Samsung introduced the first generation of 3D NAND flash memory chips. Later, 3D NAND technology continued to develop, the number of stacked layers continued to increase, and the capacity became larger and larger.
In 2021, IBM in the United States proposed the concept of "Storage-Class Memory" (SCM, Storage-Class Memory). IBM believes that SCM can replace traditional hard drives and complement DRAM.
Behind SCM is the industry's exploration of new types of memory (media).
According to the consensus of the industry, the new memory can combine the high-speed access of DRAM memory and the characteristics of NAND flash memory to retain data after the power is turned off, breaking the boundaries of memory and flash memory, making it one and achieving lower power consumption, Longer life, faster speed.
At present, there are mainly several new types of memory: phase change memory (PCM), resistive memory (ReRAM/RRAM), ferroelectric memory (FeRAM/FRAM), magnetic memory (MRAM, the second generation is STT-RAM), carbon nanometer tube memory.
To summarize, I drew a complete semiconductor storage classification diagram:
Figure. 15
In the picture above, there are many types of memory. But as I said earlier, you can focus on DRAM, NAND Flash and NOR Flash. Because, in the current market, these three kinds of memory account for more than 96% of the market share.
In fact, all memories will find their own position in the market based on their own characteristics and exert their own value.
Generally speaking, the more powerful the memory, the more expensive it will be, and the closer it will be to the computing chip (CPU/GPU, etc.). A memory with weak performance can meet the requirements of low storage latency and insensitive to writing speed, reducing costs.
Typical memory hierarchy in a computer system
In fact, the evolution of semiconductor storage technology has always benefited from Moore's Law, which has continuously improved performance while reducing costs. In the future, as Moore's Law gradually fails, where will semiconductor storage technology go, and new storage media can rise? let us wait and see.
Catalogue
FPGA / CPLD
Memory
MOS 
MCU 
DSP
OCEP
Secondary 
Other 