2003网站太阳集团(中国)官方网站

Company Trend(Feb. 4th)

SK Hynix plans to build HBM plant in USA to help make AI chips

Recently, two insiders revealed that SK Hynix, a Korean storage chip manufacturer, plans to build an advanced packaging plant in Indiana, USA to produce HBM chips which will be integrated with NVIDIAs AI GPUs to enable AI chips to be made in the USA. Clearly, this move will greatly drive the goal of the Biden administration to bring more AI chip supply chains into the USA.

According to the report, SK Hynix will build this advanced packaging plant in USA to exclusively produce stacking standard DRAM chips to finally make HBM chips, which will be integrated with Nvidias GPUs. The GUPs can be used to train large-scale generative AI models such as ChatGPT for OpenAI.

It should be noted that TSMC is currently investing in US$ 40bn in Arizona to build two advanced process fabs, of which the 4nm wafer fab of the first phase will start mass production next year. By that time, Nvidia is expected to arrange its AI GPU production at TSMCs wafer fab in Arizona. It seems that SK Hynix’s advanced packaging plant in Indiana aims to meet Nvidias localization supply demand, and also complies with Biden administrations demand for "strengthening semiconductor manufacturing in the USA", and this move is expected to be granted with subsiding funds under the CHIPS and Science Act.

Comments

Currently SK Hynix officially said that it is currently considering investing in the USA, but has not made final decision yet. However, if SK Hynix builds an advanced HBM packaging plant in the USA, together with TSMCs fab in Arizona, it obviously means that Nvidia will eventually have the capacity to further produce GPUs in the USA.

Currently, the demand for HBMs from American customers is growing constantly. At the same time, it is necessary for fabs to cooperate with Therefore, the deeper integration of SK Hynixs advanced packaging technology with Nvidias supply chain will also help the Korean company to cope with competitions from other HBM manufacturers such as Samsung and Micron.

Company Trend(Feb. 4th)

Samsung 3nm procss faces major issues in trial production tests, with yield rate 0%!

According to the report of the Korean media, Samsung’s 3nm process faces major issues with the trial production tests and the yield rate hit 0%. The report pointed out that the 3nm Exynos 2500 failed to pass the quality testing due to defects, causing the delay of mass production of the chipsets for Galaxy Watch 7.

Exynos 2500 will follow the architecture of the previous generation 10-core CPU and adopt the new Cortex-X5 core. The upgrade of Exynos 2500 lies in that it adopts Cortex-X5 and Cortex-A730 and it will show much better performance compared to Cortex-X4 and Cortex-A720 employed by Exynos 2400. Unfortunately, the clock frequencies of Cortex-X5 and Cortex-X4 have little difference and the test frequency is between 3.20GHz and 3.30GHz. According to Samsungs final decision, the frequency will be increased only by 100MHz or no difference

Samsung delayed the mass production of Exynos 2500 due to the failure in its trial production, and it is unclear whether Samsung can solve the problem of yield rate in a timely manner.

Comments

Samsung is radical in launching its 3nm process. Compared to the conservative strategy of TSMC when choosing GAA transistor for its 2nm process node, Samsung adopts GAA transistor technology for its first generation of 3nm process. Moreover, Samsung’s transistor adopts MBCFET-based node, known as SF3E (3GAE). Samsungs 3nm GAA technology employs a wider channel nano-sheet that delivers higher performance and energy-consumption ratio than GAA with narrow-channel nanowire. As for 3nm GAA technology, Samsung has the ability to adjust the channel width of nanotransistor and optimize power consumption & performance to meet diverse needs of customers.

In general, chips manufactured using more advanced process processes will feature better performance, reduce power consumption and area， resulting in lower overall costs, and new processes will give chip designers larger space to give play to their design ideas. As a result, Qualcomm, Apple, Nvidia and other large companies prefer to adopt advanced processes to make chips. However, the chips based on Samsungs 3nm process is still far from mass production for commercial purpose. This also means that the advanced process will still be monopolized by TSMC for some time, and it is needless to say such process will be profitable.

Company Trend(Feb. 5th)

Nvidia’s cut-down AI chip now available for pre-orders, priced at RMB 110,000/unit, with lower performance but similar price to Huawei’s AI chip

Recently, it is rumored that Nvidia has started taking pre-orders for a new tailored-for-China cut down AI chip from distributors who are pricing it on par with a rival product from Huawei. The H20 is the most powerful of the three graphics cards designed by Nvidia for the Chinese market. However, the H20 is less powerful than Nvidia’s flagships H100 and H800, the later China-specific card that was also banned by the Biden administration in October.

According to the news, Nvidia has in recent weeks been pricing orders for H20 distributors in China in a range of US$ 12,000 to US$ 15,000 per card, but some distributors have started advertising the chips by pricing them at about RMB 110,000 (equivalent to US$ 15,320). By comparison, Huawei’s 910B is being sold for around RMB 120,000, sources said.

At present, distributors are offering H20 servers, which are pre-configured with 8 of the AI chips, for RMB1.4mn. By comparison, servers that used 8 of the H800 chips were sold at around RMB 2mn when they were launched a year ago. Distributors told customers that they will be able to start delivering the H20 chips in small batches in Q1 2024 and in larger quantities from Q2.

Comments

In some key areas, the H20 is less powerful than Huawei’s Ascend 910B. Although the H20 has high cache and high bandwidth, its rated computing power is half of its rival’s capability. However, the H20 seems to have an advantage over the 910B in terms of interconnection speed. In fact, the H20 has more cache and bandwidth than 910B, and its bandwidth is twice that of 910B. This means that the H20 has an advantage in terms of interconnection speed, which measures how quickly data can be transferred between chips. As a result, the H20 remains competitive with the 910B in applications that require linking a large number of chips together to work as a system. This advantage is particularly evident in the context of training large models. Although Nvidia’s China-specific chips have lower performance than domestic alternative chips, their ecosystem coverage is much wider, and homemade alternative are still in short supply.  Therefore, for the next 12 months, Nvidia’s chips will remain the first choice for Chinese technology enterprises.

Company Trend(Feb. 6th)

Intel delays its US$ 20bn plant in USA

Recently, Intel announced the bad news that its new chip plant will be delayed. Intels US$ 20bn chip manufacturing project in Ohio was delayed, and it is expected to be finished until the end of 2016 rather than put into operation in 2025. The delay was due to the market and the slow roll-out of the Federal funding for the project.

An Intel spokesperson stated that the company would not be able to meet the target of starting production in 2025, as announced in January 2022 for the first time. Even if Intel is keen to build a project in the heart of Silicon Valley, the completion schedule is out of expectation, and the exact completion date depends on a number of factors. Even if Intel had a plan to start the No.1 plant in Ohio ahead of schedule and keep its construction on track, The federal government of the USA has not officially announced a major subsidy for chip plants, and it is expected to be announced at the end of March. Now Intel has not received the expected US$ 600mn subsidy appropriated by the government yet, and the project has been delayed due to labor and cost issues.

Comments

In 2022, Intel boasted that it would build the worlds largest chip plant in Ohio in the USA. Intel said that since Intel can help build a Silicon Valley, it can also build a Silicon Center. However, this grand "blueprint" is being encumbered by reality. The original commissioning date of the plant has been postponed from 2025 to 2026 when it will be fully completed. Although the slow granting of the subsidy from the government of the USA is the main reason for the decision of delay, another important reason of the delay is the challenge faced by the development of the market. Intel indicated that it is still committed to the project, but declined to comment on the exact timeline. Intels financial report of Q4 2023 and its performance outlook of Q1 2024 have showed that Intel will not improve the schedule in a short time, and the commissioning date of the plant awaits to be verified.

Domestic Trend

Sunwise Semiconductor wins RMB100mn strategic investment from Ruiyue Holding Group(Feb. 4th)

According to the news of its official Wechat account of Ruiyue Holding Group, the signing ceremony of investment agreement between Shenzhen Ruiyue Holding Group and Fujian Sunwise Semiconductor Technology Co., Ltd. was held at the conference room of Ruiyue Holding Group. It is learned that Ruiyue Holding Group strategically invested RMB 100mn in Sunwise Semiconductor to help it achieve an annual output of 750,000 units of gallium oxide epitaxial wafers and 1.2 billion units of filter chips.

According to data, Fujian Sunwise Semiconductor is a high-tech enterprise with full-chain core technology of BAW in 5G communications and it operates in the IDM mode. Sunwise Semiconductor has the core technology of manufacturing BAW chips based on the single crystalline gallium oxide as piezoelectric material. The company owns the independent intellectual property right and has achieved many global firsts and domestic first disruptive innovative technologies. The performance of the chip is 20%+ better than foreign similar products. It has passed the verification of many well-known brand customers, and can effectively remove the bottleneck in the field of 5G RF filter chips. Sunwise Semiconductor is committed to building an ecosystem of the third-and fourth-generation compounds semiconductor chip industry, and is the worlds leading manufacturer of wide bandgap semiconductor epitaxial film materials.

MZ Optoelectronics ships the first 28nm overlay precision measuring device(Feb. 6th)

According to the news of the Wechat account of Shanghai Integrated Circuit Industry Association (SICA), MZ Optoelectronics Technology (Shanghai) Co., Ltd. (MZ Optoelectronics) successfully shipped its first 28nm engraving precision measuring device at Pudong Zhangjiang, indicating that MZ Optoelectronics has become one of a few domestic enterprises with technological strength and technology reserve up to the key process nodes.

At the shipping ceremony, the secretary general of SICA pointed out that at present the IC industry chain of China faces bottlenecks in terms of engraving measuring device and there are just a few enterprises with the strength to develop such devices. Therefore, it is not easy for MZ Electronics to focus on this segment to develop products and win the trust of customers. The shipment of the device shows that the whole team of MZ Optoelectronics and its technical strength & technical reserve have been recognized by the market.

It is learned that MZ Optoelectronics Technology (Shanghai) Co., Ltd. is a company engaged in providing precision measuring and detection devices related to process control and yield rate management based on advanced optical and computer graphics image technologies for semiconductor wafer manufacturing. MZ Optoelectronics provides integrated semiconductor measuring and detection solutions of optics, machine, electronics, algorithm, computing and software based on technologies such as optical design, image processing, algorithm analysis, and motion control. The products of MZ Optoelectronics mainly include engraving precision measurement, fully automatic wafer defect detection, non-patterned wafer inspection, key size measurement, and full field non-contact mask aligner across five categories.

L&K Engineering won bid of clean room project for 12-inch memory base(Feb. 6th)

According to the news of SEMI, L&K Engineering (Suzhou) Co., Ltd. announced on the evening of February 5 that it has received the bid-winning notice from the IT Electronics Eleventh Design & Research Institute Scientific and Technological Engineering Co., Ltd. (EDRI), confirming that L&K Engineering has been accepted as the contract unit of the clean room system for the phase II project of the 12-inch memory wafer manufacturing base.

The announcement showed that the contract price under the bid was RMB 728mn, and the owner of the project is CXMT, a company mainly engaged in storage technology services, design, manufacturing and processing of IC and sales of electronic products, etc. L&K Engineering said that the smooth implementation of the project will help improve its business undertaking ability, accumulate more experience for developing subsequent projects and cooperation, and will have a positive impact on its business performance.