Known as "S HBM", Winbond CUBE Advanced Edge AI Storage

Electronic Enthusiast Network Report (Text/Huang Jingjing) Unlike AI training, GPU and HBM are used, and what kind of memory method is used for edge AI, DDR, GDDR, LPDDR, etc. A few days ago, Judy, director of electronic products at Winbond, was interviewed by media including electronics enthusiast network, sharing the application advantages of Winbond's CUBE products in edge AI and her views on the storage application market.

CUBE:小号HBM

"Winbond Electronics has been promoting CUBE products in the past two or three years, and we can see CUBE as a small HBM." Judy said that the three major factors that promote the implementation and development of AI are computing power, transportation capacity and storage capacity. Generally, the GPU of an AI server requires a large-capacity, high-bandwidth HBM, and the GPU and memory chip are combined in a single chip for system-in-package. Winbond's CUBE is a similar idea in the big picture.

Winbond's CUBE (Customized Ultra-Bandwidth Elements) is an innovative memory product designed to address the challenges of SoC (System on Chip) in DRAM packaging. This compact, ultra-high-bandwidth DRAM is designed for edge computing, and by placing the SoC die on top of the DRAM die, CUBE technology can reduce cost and size while eliminating the use of the SoC's Through-Silicon Via (TSV) process. In addition, it improves heat dissipation and is particularly suitable for applications where low power consumption, high bandwidth, and low- to medium-capacity memory are required.

CUBE has the following benefits:

• Power Savings: CUBE delivers superior power efficiency with less than 1pJ/bit to ensure extended uptime and optimized energy usage.
• Superior Performance: With bandwidth from 16GB/s to 256GB/s, CUBE delivers performance gains that are well above industry standards.
• Smaller size: The CUBE has a smaller form factor. It is currently based on the 20nm standard, which can provide 1Gb-8Gb capacity per chip, and there will be a 16nm standard in 2025. The introduction of through-silicon vias (TSVs) further enhances performance with improved signal integrity, power integrity, reduced IO area at 9 um pitch, and better heat dissipation (CUBE down, SoC up).
• Cost-effective and high-bandwidth: CUBE's IO speeds can reach up to 2Gbps at 1K IO, and when integrated with mature SoCs such as 28nm and 22nm, CUBE can reach 32GB/s to 256GB/s bandwidth, which is equivalent to HBM2 bandwidth, and is equivalent to 4 to 32 LP-DDR4x 4266Mbps x16 IOs.
• SoC chip size reduction: SoCs (without TSV, top) are stacked on top of the CUBE. If the TSV area loss is removed, its chip size may be smaller. It can bring more significant cost advantages to edge AI devices.

Judy said that the biggest difference between CUBE and HBM is the capacity. The capacity of HBM is very large, and the capacity of CUBE is a little smaller. Because Winbond mainly engages in niche storage, small and medium-capacity storage, the application market it aims at is mainly at the edge, not HBM's computing for the cloud and server. And the current general trend is the sinking of computing power, that is, the computing power of the cloud is delegated to the edge, because the edge side responds faster and is more secure. In the cloud, many customers are worried about data security and privacy protection. In addition, inference or a small amount of training can also be done at the edge, and the strong AI capabilities at the edge can also reduce costs.

Taking automotive applications as an example, there are many sensors on the car, including radars and cameras, and if the sensing computing power of the device-side sensors is strong enough and the data is preprocessed, the main control chip is not required for processing. This is a typical way to reduce the cost of edge-side AI processing.

At present, the potential of CUBE in the application of independent ISP chips is very high. Judy said that now there are more and more cameras, and when the ISP of the main control processor is not enough, manufacturers will develop their own ISP chips, which brings great opportunities for low-power LPDDR4 or CUBE. In addition, IP Cameras for home or industry use, DDR4 or LPDDR4 are currently used for the most high-end storage, but the application opportunities of CUBE are not ruled out in the future.

Car storage, keeping pace with the times

According to the data, in 2022, Winbond Electronics is the world's fifth largest automotive storage manufacturer in terms of revenue, and Winbond Electronics provides a very complete range of small and medium-capacity storage products, including NOR Flash, DDR3, etc., which are widely used in automotive sensors, ADAS, intelligent cockpits, etc.

Judy said that the current dashboard is integrated, with one core and multiple screens, that is, integrating entertainment screens, information screens and instrument screens. If the main control chip is broken for some reason, the screen can still be displayed, which depends on the small system inside, and the small system can be matched with Winbond's small and medium-capacity storage.

As a storage manufacturer, Winbond Electronics used to be a Tier1 direct user in the automotive field, and now it will also supply directly to car manufacturers, and many IDH or main control platform manufacturers also interact closely with us, and Huabang will adapt to their solutions to provide customers. This is a change in the automotive ecology, and Winbond is also keeping pace with the times.

In addition, the competition in the automotive industry is very fierce, the product iteration cycle is shorter, some manufacturers launch one or even several new cars a year, and the car is also facing the current situation of rapid price reduction and market grabbing, so the product iteration and supply capacity of memory chip manufacturers will bring new challenges. Winbond is also reducing costs by improving R&D capabilities and manufacturing processes to help customers achieve the goal of reducing costs.

Next year or a big year for storage, there is a big opportunity for small and medium-capacity storage

The increase of small and medium-capacity memory chips is smaller than that of large-capacity, and the price increase trend is slower. Judy believes that the top companies in the storage industry have been losing money before, which is an abnormal industry phenomenon, so the subsequent price modification can promote the healthy and sustainable development of the entire industry.

For the market trend from the second half of this year to the first half of next year, Winbond remains optimistic on the demand side. In particular, the increase in shipments of computers and smartphones in the electronics industry will drive the entire industry up.

In addition, HBM accounts for single-digit total DRAM production capacity this year, and according to third-party reports, HBM may account for 30% of production capacity next year. Coupled with the fact that PCs are moving from DDR4 to DDR5, it will also take up capacity. Judy expects that next year may be a big year for storage, especially in the small, medium, and medium-capacity, niche storage market. As a result, Winbond continued to expand its production capacity to cope with the increase in demand in the future, despite the fact that the market was not very good last year.

brief summary

For edge AI, in addition to CUBE products, Winbond focuses on expanding some customized and semi-customized innovative products. In fact, customer needs vary greatly, especially since some leading manufacturers have their own design requirements, and Winbond will continue to provide innovative storage to meet customer needs.