Samsung unveils UFS 5.0 amid sharp selloff in Korean AI stocks

Samsung Electronics unveiled the successful development of its Universal Flash Storage 5.0 (UFS 5.0), offering sequential read speeds up to 10.8 GB/s, the fastest recorded in this standard.

The product is being released amid increased AI workload migration from cloud data centers to edge devices, making it increasingly necessary for the storage systems to become a potential performance bottleneck for on-device inference.

This comes in when the KOSPI Index of South Korea experienced an almost 10% drop on Tuesday due to the selling off of key semiconductor and technology stocks amid weak US tech stocks. The index declined by 910.71 points, representing a fall of 9.99% to end at 8,203.84 having initially risen above 9,175 levels in volatile trading.

It was the fourth time this year and the tenth time ever that trading was suspended in the market. Key firms recorded significant declines in the market, including Samsung Electronics by 12.31%, SK Hynix by 12.47%, Hyundai Motor by 12.05% and LG Energy Solution by 6.1%.

Samsung speeds up local AI

According to Samsung’s report, the sequential read and write speeds offered by UFS 5.0 are twice those offered by UFS 4.1 available in today’s flagship smartphones. In particular, sequential write performance can hit 9.5 GB/s.

The throughput of storage systems is important for large language models because when run locally, these models have to be loaded from flash to DRAM for inference, making bandwidth an important limitation.

This can be explained with the help of a simple numerical illustration. Loading the 7-billion parameter model which otherwise requires 4-5 GB of space in a 4-bit configuration, will take 0.98 seconds with a speed of 4.6 GB per second in the case of UFS 4.1. On the other hand, it will take 0.42 seconds for UFS 5.0 to do the job. This marks a decrease in time by about 57%, without even accounting for any system overhead or thermal limitations.

Storage does the heavy lifting

Performance of on-device AI is increasingly being determined by data movement through memory rather than just computational capabilities alone.

Usually, during inference, model parameters are stored in UFS storage, loaded into the DRAM, and computed by the system-on-chip (SoC). With the increasing size of models, the storage bandwidth limits the responsiveness of the system.

Samsung reaches the maximum throughput of the HS-GEAR6 protocol, which was introduced as part of MIPI M-PHY 6.0 and incorporated into the UFS 5.0 stack. The interface provides up to 46.6 Gbps for each lane, meaning 10.8 GB/s per dual-lane interface.

Participants in the industry seem to be on track for the same performance. KIOXIA started shipment of 512GB UFS 5.0 samples and prepares to launch 1TB models, indicating that the mobile storage sector is slowly adopting a common platform that offers faster and AI-capable memory solutions.

Semiconductor rally meets a sharp pullback in Korean equities

This comes at a time when South Korea’s semiconductor-laden stock market is experiencing sharp swings. Specifically, on June 23, the KOSPI experienced an overall selloff led by tech stocks, where Samsung Electronics and SK Hynix fell double-digits as foreign funds sold off crowded AI trades.

It is following a robust rally for several months in semiconductors driven by AI demand. The two tech giants had reached a valuation of more than $1 trillion each, driven by strong prices of memory used in AI accelerator chips and data center hardware.

SK Hynix had even managed to beat Samsung Electronics to become South Korea’s most valuable listed firm due to its exposure to the HBM production chain to Nvidia and hyperscale cloud providers.

This illustrates how the KOSPI index has become very concentrated, especially on semiconductor stocks, whose weight in the index increases the effect of rapid changes in positioning in AI trades globally. 

Foreign flows and valuation pressures affect Korean chip stocks

There has been reduced foreign investor interest in technology stocks in South Korea in recent trading sessions, with the selloff happening particularly in large semiconductor firms.

The two largest semiconductor firms, Samsung Electronics, and SK Hynix, are currently seeing their share prices fall in tandem with the overall market index as rebalancing occurs after the surge caused by AI technology.

The correction in prices is seen not as a result of any specific trigger but rather a reduction in a speculative hot AI trade for underperformers which valuation looks good.

Both the increase in prices of semiconductor stocks as well as their drop can be largely attributed to foreign inflows. The previously heavy selling in semiconductor firms led to large intraday moves in the KOSPI, especially when the sentiment in the technology stocks was negative globally.

Nonetheless, despite the recent move in prices, the fundamentals of the semiconductor industry are in sync with those required by AI infrastructure, with SK Hynix benefiting from its capabilities in high-bandwidth memory used in AI training, while Samsung has exposure to DRAM, NAND, and packaging.

At the same time, sentiment has turned cautious after the protracted rise in AI shares, with the upside in valuations now being questioned by market participants. The macro conditions, which feature an increase in bond yields in the U.S., as well as a rotation away from growth shares, have further added to the pressures, resulting in a simultaneous drop in Asia-based semiconductor indices.

If you're reading this, you’re already ahead. Stay there with our newsletter.