Significant Investment
Samsung Electronics plans to supply essential components for graphics processing units (GPUs), the high-bandwidth memory (HBM), and advanced packaging services to U.S.-based semiconductor company Nvidia.
According to semiconductor industry sources on Aug. 1, Samsung Electronics is currently working with Nvidia on technical verification tasks for the HBM3 for GPUs and advanced packaging services. As soon as the technical verification procedures are completed, Samsung will supply HBM3 to Nvidia and is expected to take charge of the advanced packaging that processes individual GPU chips and HBM3 into a high-performance GPU, the H100.
Previously, Nvidia entrusted most of its GPU advanced packaging volumes to TSMC. TSMC produced Nvidia’s H100 by packaging SK hynix’s HBM3 on individual GPU chips manufactured through its own process. However, with the rapid increase in H100 demand due to the recent spread of generative artificial intelligence, TSMC struggled to handle all of Nvidia’s orders. Major clients such as Microsoft have announced service disruptions due to GPU shortages, prompting Nvidia to turn to Samsung Electronics, which has the capacity for both HBM3 and advanced packaging.
Samsung Electronics plans to enhance its “one-stop service” that is responsible for all semiconductor processes such as memory semiconductors, contract semiconductor manufacturing, and advanced packaging, using the deal with Nvidia as a stepping stone. Professor Park Jae-geun of the Department of Convergence Electronics Engineering at Hanyang University stated, “The strength of Samsung Electronics is that it is a comprehensive semiconductor company with a diverse portfolio,” and added, “More companies looking for AI semiconductor production will turn to Samsung.”
Since the advent of semiconductors in the 1960s, the measure of technological competitiveness has been “how well a single chip is made.” Samsung Electronics, Intel, and TSMC have established an unbeatable position in the DRAM, CPU, and foundry sectors, respectively, based on their manufacturing competitiveness.
Recently, the situation is changing. With the proliferation of AI technology and the growing demand for high-capacity, high-performance semiconductors, “advanced packaging” technology, which maximizes performance by skillfully combining multiple chips, has emerged at the forefront. Within the semiconductor industry, it is predicted that “advanced packaging competitiveness will determine the fate of semiconductor companies in the future.”
According to the semiconductor industry on Aug. 1, there is currently fierce competition in advanced packaging technology and investment between major semiconductor companies such as Samsung Electronics, Intel, and TSMC with regard to the market related to high-bandwidth memory (HBM), a term denoting high-performance DRAM. According to market research firm Yole Intelligence, the advanced packaging market, which was worth US$37.4 billion in 2021, is expected to grow to US$65 billion by 2027.
The rise of advanced packaging began with the proliferation of generative AI like ChatGPT. In order to advance generative AI technology, it is crucial to quickly learn and serve large-scale data through semiconductors. This is why Nvidia’s GPU, which is optimized for parallel computing to process large volumes of data simultaneously, and DRAM, a storage device that communicates and exchanges data with the GPU in real-time, have become the focus.
Until now, semiconductor companies have competed to create ultra-small, low-power, high-performance chips by reducing the line width (the width of the circuit) down to the nanometer scale. However, as the line width of the latest process entered below 3 nm, the cost of miniaturizing the chips has increased, and additional technological development has become difficult.
In this context, advanced packaging emerged as an alternative. Stacking and bundling multiple chips can deliver performance comparable to operating a single high-performance chip. Recently, semiconductor companies have focused on “2.5D packaging,” which reduces potential data bottlenecks by stacking 4, 8, or 12 DRAM modules vertically to create HBM, which is more than ten times faster than regular DRAM, and placing HBM as close as possible to the GPU.
Currently, TSMC, a foundry company, is leading the advanced packaging race in the HBM and GPU sectors. The company has developed the “CoWoS,” a 2.5D package, over five generations for ten years from 2011 to 2021. On July 25, TSMC announced that it would invest 90 billion New Taiwan dollars to construct an advanced packaging factory in northern Taiwan.
Samsung Electronics has also rolled up its sleeves to nurture advanced packaging. It unveiled a 2.5D packaging technology called “X-Cube” in 2021 and established an Advanced Packaging (AVP) team at the end of last year. It is known that starting from the second quarter of next year, the company will mass-produce the “X-Cube4,” which places four HBMs with the GPU, and “X-Cube8,” which places eight HBMs, in the third quarter. Samsung is expected to supply Nvidia with the third-generation HBM, or HBM3, by the end of this year and to use X-Cube4 to package it with individual GPU chips. Intel is also proactive, planning to invest US$4.75 billion in advanced packaging facilities alone.