Semi-OT: Samsung compensating for godawful yields on HBM3

HBM3 is going to be a major battlefield, I think. Had no idea Samsung had this kind of yield issues. Not sure how having their own packaging will get them an advantage; seems like the key move will be fixing the yield issue.

A key technology used to make artificial intelligence chips is the reason the world’s biggest memory manufacturer, Samsung Electronics of Korea, has fallen behind rivals from its home country and the U.S., according to sources quoted by Reuters. Along with those designed by Wall Street’s A.I. darling NVIDIA corporation, high bandwidth memory (HBM) chips are an equally important component of the A.I. revolution that has become the center of stock market attention as the technology and industrial sectors start to expand their computing and data processing facilities to include artificial intelligence workloads.

Reuters’ sources claim that Samsung is buying equipment that will help it improve chip yields, which have often proven to be Samsung’s Achilles heel in the application processor segment of the global semiconductor industry.

Samsung Investing In Advanced High Bandwidth Memory Packaging Technology In Bid To Win NVIDIA’s A.I. Orders

Early in the A.I. race, even as NVIDIA’s shares raced to record high, semiconductor analysts wondered whether there was sufficient chip packaging capacity in the industry to meet booming demand for the firm’s GPUs that power A.I. workloads.

These constraints appear to have shifted the tide in the memory chip industry as well, as today’s report quotes five sources to share that Samsung is eager to catch up to rivals when it comes to yields. The firm’s chip production yields for HBM3 chips sit at a weak 10% to 20%, according to analysts quoted by the publication, while its Korean rival SK Hynix achieves as much as 70%. Yields are a crucial part of chip fabrication as they determine the number of usable chips in a silicon wafer.

**Low yields are a key reason behind Samsung lagging its memory players in winning NVIDIA’s HBM3 orders, believe the sources. **To make up for this, Samsung is allegedly procuring machines and materials used by chip manufacturers to fill gaps between layers within a memory chip’s layers with epoxy. Samsung relies on its own technology for HBM3 packaging and has resisted attempts to shift to the new technology called MR-MUF (Mass Reflow Molded Underfill).

SK hynix’s HBM Technology Stacking Roadmap. Image: SK hynix IEDM 2023 presentation. Sourced from:

The equipment is used in the latter process of semiconductor manufacturing, which is called packaging. Chip fabrication involves printing circuits on silicon, and the final product is then packaged to make it suitable for use in a computer. HBM3 chips sold by SK Hynix and Micron to NVIDIA work in tandem with a GPU and they are indispensable for any A.I. system.

Potential A.I. demand and existing industry investments have catalyzed semiconductor stocks this year and in 2023 despite a major glut that has plagued the consumer end of the industry. A demand supply mismatch after the coronavirus pandemic led to excessive orders from firms such as NVIDIA and AMD, and as the market cooled, the excess inventory meant that chip manufacturers such as the Taiwan Semiconductor Manufacturing Company (TSMC) saw a drastic reduction in revenue.