Besides the Middle East, Nvidia has also caused a plunge in the South Korean stock market.

```

In the past two days, Korea’s benchmark KOSPI index has dropped more than 10% each day, marking the biggest two-day decline since 2008.

The market generally believes that global risk aversion caused by Trump’s military action against Iran led to brutal declines across Asian stock markets, including Korea. However, the latest analysis points out that Nvidia has also “contributed” to Korea’s sharp fall.

A technical rumor about Nvidia has precisely hit Korean memory stocks. According to Korean analyst Jukan at Citrini7 quoting independent agency KIS, reports say Nvidia is developing a new inference chip using Groq’s on-chip SRAM architecture and plans to announce it at the GTC conference in March.

This news led to weakness in Korean memory stocks, with investors worrying that the use of SRAM would reduce demand for main memory, including HBM.

However, Korea’s stock market saw a strong rebound today. The latest market data shows that Korea’s KOSPI index surged by 11% today, tech giant Samsung Electronics soared 13%, and SK Hynix skyrocketed 15%.

SRAM inference chips impact HBM, DRAM? It may be a misjudgment

However, the market may have misjudged the impact of SRAM inference chips.

KIS made it clear: “The view that the emergence of ‘low-cost’ SRAM inference chips will reduce the use of main memory such as HBM reflects poor understanding of memory.”

From a physical perspective, SRAM units occupy larger area, and their density is lower than DRAM, making their per-bit cost significantly higher. For the same capacity, SRAM typically requires 5 to 10 times the die area of DRAM. Therefore, historically, SRAM has been used for cache or on-chip buffer applications that require ultra-low latency, rather than for main memory storing large amounts of data.

SRAM may promote diversification of memory hierarchy

SRAM architecture is not a substitute for DRAM, but rather an independent option. Compared to DRAM, architectures centered on SRAM have much lower access latency and minimize data movement.

KIS analysis states that Nvidia’s plan to utilize the Groq architecture is aimed at optimizing specific inference workloads that GPUs struggle to handle. The adoption of SRAM architecture should be understood as a unique choice for certain data center workloads requiring ultra-low latency and for real-time-responsive AI edge applications (such as robotics and autonomous driving). In fact, OpenAI has already deployed Cerebras’ SRAM chips in its data centers, and inference services built on these chips charge higher API fees than standard GPU inference services.

As the AI industry advances, the popularization of Groq-based SRAM architectures will further segment the memory hierarchy within AI infrastructure. HBM and DRAM will continue to be the main memory for large-scale model training and general inference servers. KIS concludes: “A multi-tiered memory hierarchy covering SRAM, HBM, and DRAM will become increasingly layered, ultimately driving the expansion of the total addressable market (TAM) for the entire memory industry.”

Risk Warning and DisclaimerThe market has risks; investments should be made prudently. This article does not constitute personal investment advice and does not take into account specific users’ special investment objectives, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Investment based on this information is at your own risk. ```