SK Hynix presented’ULM’ and’AIM’ as future memory semiconductor technologies that can reduce carbon emissions at Semicon Korea, the largest material, parts and equipment exhibition in Korea, which opened on the 3rd.
ULM (Ultra Low Power Memory) stands for ultra low power memory, and AIM (Artificial intelligence in Memory) stands for intelligent memory. Both are technologies that significantly improve the work efficiency of memories that store and process data, and are expected to attract attention in the post-corona era when non-face-to-face economic activity is increasing.
SK Hynix Vice President Cha Sun-yong said, “ULM is a product designed to be applied to specific applications that are more optimized for System-on-Chip (SoC). It is possible to increase capacity by ~1000 times, and in terms of energy efficiency, it can improve up to 4~5 times compared to high bandwidth memory (HBM).”
Also, “AIM is a concept that puts computational functions into memory as much power and time are consumed in a situation in which data is transmitted, and the speed can be increased by that amount.” “The fastest combination currently is a combination of a graphic processing unit (GPU) and HBM. , It is possible to improve the bandwidth by 10 times and increase the power efficiency by 7 times In addition to shortening the distance between the central processing unit (CPU) and the SoC and the DRAM chipset, it also brings the computational functions used in the existing CPU to the memory. It comes in that form,” he added.
In addition, “HBM, ULM, and AIM products all show a great improvement in terms of performance and power (power) depending on how close they are to SoC and how they are fused. It is a plan, and SK Hynix is also developing products of various solutions in addition to the existing standard products, and various collaborations are in progress for this.”
The impact and vision of ESG (environmental, social, and governance) management being promoted by SK Group on the memory semiconductor industry were also introduced.
Vice President Cha Sun-yong said, “In the future, the development of information and communication technology (ICT) technology through artificial intelligence (AI) and 5G as well as the untact industry will be greatly accelerated, and various environments ranging from smart cars to smart cities will be established.” These are all data-driven changes, and the number of data centers is expected to double within the next five years, resulting in 20 million gigabytes for DRAM and 750 million gigabytes for NAND. It is expected that a large amount of capacity will be used.”
“Environmental problems caused by the explosion of data are becoming more and more serious, and the development of ICT technology is also not free from these problems.” “I think the semiconductor industry should also strengthen its social responsibility to preserve the environment through energy savings. This is possible. In order to do this, overall cooperation in each field such as academia, semiconductors, equipment, SoC, memory, software, etc. is required. SK Hynik will actively cooperate with this.”
The following is the full text of Semicon Korea 2021 keynote speech by SK Hynix Vice President Cha Sun-Yong.
■ Sun-Yong Cha, Vice President of SK Hynix
The untact environment caused by Corona 19 is causing many changes in our daily life, life, and technology direction. According to Forbes, the main nine areas that are expected to be required in the post-pandemic and new normal era are predictions that digital transformation will accelerate further by activities in most untouched situations.
These changes were the same in a survey of C-level technology leaders in 83 countries around the world. 47% of respondents said that Corona 19 will further accelerate digital transformation and adoption of new technologies.
Corona has changed a lot, and the changed part is expected to remain in the post-corona era.
In the case of video conferences that exchange video data in real time, 1 to 3 gigabytes of data are used for an hour based on 5G mobile. Due to Corona 19, video conferences are frequently held inside SK Hynix, and many companies will be the same.
This situation means a significant increase in data usage. In the future, the development of information and communication technology (ICT) technology through artificial intelligence (AI) and 5G as well as the untact industry will be greatly accelerated, and various environments ranging from smart cars to smart cities will be established. And these are all data-driven changes.
5G transforms our society into a hyper-connected society.In particular, in the case of autonomous vehicles, the accumulated data other than the various sensors installed here is connected to the data center through the Internet, and approximately 4000 gigabytes of data are used per day. Is expected.
Accordingly, data usage is expected to more than double in the next five years. With various streaming services, vehicle-to-vehicle (V2V), person-to-person (M2M) communication, and other fast data sharing and real-time data that disappears immediately, the data growth will be even steeper. It can be said to be an explosion of data.
Data is basically stored in a memory device. A representative one is the data center. The number of data centers is growing by 15% annually. A very high growth rate is expected. Growth is expected to double within the next five years. Estimating the amount of memory required per data center is 20 million gigabytes for DRAM. NAND is expected to use as much capacity as 750 million gigabytes.
Eventually, there will be a situation where the amount of data increase cannot match the amount of memory supply. In other words, there will be insufficient memory to store data. Memory will expand a lot of roles and applications in the ICT environment. In order to implement the Internet of Things (IoT), data generated by edge devices will go through several processes to form big data, which will be stored in the cloud. This will be re-analyzed and processed to be born as new information, and it will form a virtuous cycle ecosystem that improves the quality of life by reaching edge devices.
The role of memory that stores and processes information is essential. There was a supercycle of the memory market in 2018. In terms of memory supply, we did not predict that data center demand would increase so rapidly. So there was a lot of shortage.
Various organizations are predicting that the supercycle will arrive from this year. It can be said that it is an increase in memory usage due to the increase in data based on 5G and AI and the spread of the untouch environment due to Corona 19. SK Hynix, which supplies both DRAM and NAND, will develop and provide memory with very good performance due to data growth.
Environmental problems caused by data explosion are also important. Gradually, environmental problems are becoming very serious. The development of ICT technology is also not free from this problem. Currently, video data accounts for about 80% of Internet traffic. Some predict that by 2030, video streaming alone will consume 4.1% of the world’s power. According to the results of a study by an environmental group, 1.6 kilograms of carbon dioxide are produced when playing video online for 30 minutes. This is the same amount that occurs when a car drives a distance of 6.3 kilometers. Greenpeace estimates that this year’s energy consumption for storing data online such as search and cloud is 10 trillion kilowatts per year. This is four times the amount of electricity used per year in Korea.
A huge amount of energy is being used to transmit data. During this explosive increase in data, power consumption and carbon dioxide are increasing. From the standpoint of developing a memory device, we cannot but consider this. In addition to economic value, it is thought that we must play a role in social value.
SK Groups, including SK Hynix, are practicing the management philosophy of Double Bottom Line (DBL). It is recognized as essential for the continuity of the business. In addition, six affiliates, including SK Hynix, have joined RE100 for the first time in Korea. By 2050, it has declared that it will deal with all the power required in the semiconductor manufacturing process with renewable energy.
When looking at the memory required for each application, it is largely organized into three types. In order to store a lot of data, the capacity must be large, the data processing speed must be fast, and the power consumption must be low.
These requirements have been and will remain the same in the past. The demand level will be higher and higher for the use of 5G and AI. In terms of energy efficiency, all of these areas can be said to be in the same direction considering the energy consumption per unit bit and per unit performance.
It can be said that semiconductor technology, including memory, has evolved in this direction to increase energy efficiency. Looking at how much energy efficiency has been contributed through the advancement of memory technology over the past 30 years, it has reduced energy consumption to the level that can travel the earth twice by 1 liter, which is similar to a car.
The advancement of this technology is achieved through miniaturization. The same is true for DRAM and NAND. In memory technology, both DRAM and NAND are progressing through miniaturization respectively.
DRAM is currently mass-producing 10 nanometer 3G 1z products, and the 4G will complete development within the first half of this year. NAND developed and introduced a new technology called 4D in 96-layer, and is currently mass-producing 128-layer products.
Refinement and stacking (3D, 4D) are the basis of technological advancement. DRAM has continuously changed in terms of product type through miniaturization. In the case of graphics and mobile products, computing bandwidth has increased and power consumption has evolved in a direction that decreases. I think these changes will continue to evolve.
What effect does micronization and stacking technology have in terms of energy efficiency? Similar to both DRAM and NAND, energy savings of about 10% were possible for each generation. In addition to saving energy through miniaturization, it was also possible to improve performance and reduce power consumption through product changes. For example, looking at DRAM computing products, energy savings of more than 20% were achieved by changing to DDR2, DDR3, and DDR5. As mobile evolves to LPDDR2, LPDDR3, LPDDR4, and LPDDR5, it is also realizing greater energy savings. In the case of SSD, compared to the existing HDD, NAND also saved a lot of energy. This is the direction to reduce carbon dioxide emissions.
By converting a hard disk to an SSD or a DRAM type from DDR4 to DDR5 in the data center, a total of 4 terawatt hours (TWh) of power consumption can be reduced. This power consumption is a tremendous amount of energy that all streetlights in Korea can be turned on for a year or a Korean home can use it for a month.
Process refinement and lamination should continue, but unfortunately, the difficulty of the technology is increasing even further. The equipment industry and academia are working together to solve this problem, and various ways are being sought to overcome technological limitations. In order to obtain the level of performance and power required in the ICT environment, a new approach is also required.
The demand for memory performance and capacity is increasing. In terms of performance, system power consumption is considerably higher in accessing DRAM than in operation process or accessing S-RAM. This means that memory speed and operating performance can determine system settings.
The traditional memory system (Hierarchy) consists of cache, main memory (DRAM), and storage (HDD). The cache is fast but the capacity is small, and the HDD is large but the speed is slow. Still, many systems have this look. In terms of the performance of the system, the difference in performance of each is large, and these are being subdivided recently. SSDs, for example, are becoming fast storage devices.
In order to fill the gap between DRAM and cache, the need for bandwidth memory is emerging. In addition to the traditional method for improving performance in the existing structure, innovative memory products are being developed to overcome this.
First of all, it is high bandwidth memory (HBM). General memories are manufactured as modules. In addition, it is mounted in a socket and is separated from the central processing unit (CPU). Or, it is used in package form like mobile. In this case, from the point of view of the CPU that uses the data stored in the memory, the memory is far away and consumes a lot of power during operation. HBM is a structure in which DRAM chipsets are stacked through a silicon through-electrode (TSV). The signal from this is adjusted by the logic die located at the bottom to exchange data with the CPU through the interposer. This itself consists of one package die. This package type is called 2.5D. This means that the distance between CPUs or graphics processing units (GPUs) is getting closer.
In addition, HBM employs 1024 inputs/outputs (I/Os) due to its structure. 1024 data goes back and forth at once. This is an increase of approximately 32 times compared to normal DRAM I/O (up to 32).
Bandwis refers to the data transmission rate per unit time. For example, it is likened to the traffic volume of vehicles on the road, meaning how much traffic can be carried. To widen this, there are ways to speed up the vehicle or widen the road. HBM widens the road from 32 lanes to 1024 lanes.
HBM is a product that can improve energy efficiency by about 40% in terms of power consumption. It is evolving into HBM1, HBM2, and HBM3. Energy efficiency is also continuously improving. With these characteristics, it is being spotlighted as an AI solution.
There is also ULM, an ultra-low-power memory. It has a system-on-chip (SoC) and a 2.5D package structure. It is a product designed to be applied to specific applications that are more optimized for SoC. It is a custom memory that is not a general purpose, and in the case of general purpose, it performs various functions and satisfies a given specification. Also, you have to design with a fairly large margin. On the other hand, custom products can eliminate blocks that are not needed by design to support specific functions. So, you can design products efficiently.
This product is designed to replace S-RAM. The capacity can be increased by 100 to 1000 times compared to S-RAM, and in terms of energy efficiency, it can be improved by 4 to 5 times compared to HBM.
HBM was also a structure closer to SoC through 2.5D structure, but ULM is a product that can get more effects by designing it closer to a specific customized SoC.
These things can lead to more performance improvements and more power savings through collaboration with SoC.
There is also AIM. Recently, there are many efforts to improve data processing speed by using a GPU other than the existing CPU for AI applications or by combining it with an AI accelerator such as a field programmable gate array (FPGA). Unfortunately, however, there are still obstacles to the ability to transfer data between AI and memory in the current architecture. So, the concept of process in memory (PIM) is being discussed and development is underway. Previously, the computation function was only in the CPU and GPU, and the memory only transferred data. AIM is a concept that puts computational functions into memory as much as power and time are consumed in a situation in which data is transmitted. That means that the speed can be faster.
The fastest combination in this case is currently the GPU and HBM combination. It is possible to improve 10 times the bandwidth and increase power efficiency by 7 times. It is a form that not only shortens the distance between the CPU and SoC and the DRAM chipset, but also brings the computational functions used in the existing CPU into memory.
HBM, ULM, and AIM products all show a great improvement in performance and power depending on how close they are to SoC and how they are fused. We plan to expand this portfolio. This trend is the same concept as the convergence trend of not only memory devices but also various ICT devices around them.
In computing, the basic concept of von Neumann computing is that memory for storing data and CPU for calculating data are separated, whereas these areas are gradually becoming converged, and ultimately, the era of data storage and computation will come at the same time. .
Related Articles
Samsung, SK, and LG see a’super gap’ again this year
‘Semicon Korea’, Korea’s largest general manager exhibition, held on February 3rd
SK Hynix completes EUV production base’M16’…
SK hynix “This year’s business performance will improve from the previous year”
SK hynix is also developing products of various solutions other than the existing standard products, and is conducting various collaborations for this. It is also expected to expand further.
The semiconductor industry should also strengthen its social responsibility for environmental conservation through energy savings. To make this possible, I think that overall cooperation in each field such as academia, semiconductors, equipment, SoC, memory, and software is necessary. SK Hynik will actively cooperate with this.