'Father of HBM' Kim Joungho: "TurboQuant Not Disruptive Enough to Shake Memory Demand"

Expert Analysis of TurboQuant's Impact from Academia, Industry, and AI Companies

"Software and Hardware Semiconductor Innovation Will Continue"

"Verification Needed... Memory Demand Will Not Decrease"

"AI Research Costs Expected to Fall"

Google logo. Reuters Yonhap News

Despite market concerns that Google's artificial intelligence (AI) compression algorithm 'TurboQuant' could reduce demand for Korean memory semiconductors, experts from various fields have analyzed that memory demand will remain strong, and that the advent of such technology will actually present significant opportunities for AI-driven research.

Kim Joungho, Professor at KAIST

Kim Joungho, a professor at KAIST known as the 'father of high-bandwidth memory (HBM),' welcomed the emergence of TurboQuant in a phone interview with The Asia Business Daily on March 27, but stated, "TurboQuant is essentially a continuation of existing compression technologies or quantization methods. While it does introduce a new direction, there is no need to overinterpret its impact." According to Google, TurboQuant is an algorithm that compresses the KV cache of AI models to 3 bits, reducing memory usage by six times and increasing speed by eight times.

Professor Kim also pointed out the technical limitations of TurboQuant. He explained, "Compressing the KV cache and then decompressing it for use inevitably causes additional latency," and added, "This process can also lead to issues like signal loss or hallucinations, so thorough validation is necessary."

He further commented, "When the context length increases or extends to multimodal applications, it is difficult to guarantee effectiveness," and emphasized, "Verifying stability in real-world service environments is crucial."

Professor Kim drew a clear line regarding market impact as well. He predicted, "This technology is unlikely to be impactful enough to fundamentally disrupt the structure of memory demand," and added, "Its effect on demand for high-performance memory such as HBM will also be limited." He did note, however, that NAND flash memory used in storage devices could be affected.

Regarding the decline in stock prices of memory semiconductor companies like Samsung Electronics, SK hynix, Micron, and SanDisk on Korean, US, and Japanese stock markets, Professor Kim assessed this as a temporary event. He remarked, "The market reaction has been somewhat overblown, and it appears similar to the trend observed when DeepSeek emerged last year."

Professor Kim concluded, "Efforts to overcome the memory barrier in AI infrastructure will continue," and predicted, "Innovations that combine software algorithms and semiconductor architecture are likely to persist going forward."

Jinwon Lee, Chief Technology Officer (CTO) of HyperAccel

Jinwon Lee, Chief Technology Officer (CTO) of HyperAccel, a neural processing unit (NPU) developer, also expressed a cautious perspective. He said, "KV compression itself is an area with extensive existing research, and TurboQuant can be viewed as an extension of current quantization technologies." He further explained, "The KV cache is stored immediately upon generation and is used directly in computation, so additional operations are required during compression and decompression. If this process is not optimized, memory usage may decrease but speed could actually decline."

He also presented a different view from the market consensus regarding the demand side. He predicted, "If memory usage decreases, there will actually be increased demand for using longer contexts, and as AI utilization expands overall, total memory demand is likely to rise." He referred to the 'Jevons Paradox,' which states that as resource use efficiency improves and costs fall, total consumption of that resource can explode as a result.

CTO Lee added, "Given that agentic AI generates an overwhelming amount of KV cache, there is absolutely no chance that demand for memory semiconductors will decrease going forward."

Taehyung Kim, CEO of Bionexerus

In research fields that directly utilize AI, there is a sense of anticipation. Taehyung Kim, CEO of Bionexerus, who is conducting scientific research using AI agents, said, "AI scientists face a significant burden from having to process many tokens to read hundreds of thousands of papers and formulate hypotheses, which has led to much concern about optimization. Google has now provided an alternative." He continued, "In bio AI, where we need to read and connect papers, patents, omics data, and analysis results together, memory efficiency and search costs are critical factors for success." He expressed hope that TurboQuant will alleviate some of the burdens associated with bio AI development.

This content was produced with the assistance of AI translation services.