According to industry sources, NVIDIA's quality testing poses a challenge to memory manufacturers, because compared with traditional DRAM products, HBM's yield is significantly lower.
Companies such as TSMC and Samsung Foundry have previously faced the challenge of maintaining optimal yields when processing single silicon wafers, but the problem has now spread to the HBM industry.
According to the news, storage manufacturers such as Micron and SK hynix will compete in the qualification test of NVIDIA's next-generation AI GPU, and it seems that the gap is not large, and the yield will be an obstacle for manufacturers.
It is reported that in the HBM manufacturing process, the complexity of multi-layer stacking leads to low yields, and the chiplets are connected through the through-silicon via (TSV) process, which increases the chance of defects in the manufacturing process. If one layer of HBM proves to be defective, the entire stack is discarded, making it difficult to improve yield.
According to the source market, the overall yield rate of HBM memory chips is about 65%, among which Micron and SK hynix seem to be in a leading position in this competition. It is reported that Micron has started to produce HBM3E memory chips for NVIDIA's latest H200 AI GPU, as it has passed the certification stage set by Team Green.
In an official announcement on February 21, SK hynix Vice President Kim Ki-tae pointed out that although external instability still exists, the memory chip market is expected to gradually recover this year, and applications such as PCs and smartphones will not only increase the sales of HBM3E, but also drive the demand for products such as DDR5 and LPDDR5T. The executive said that the company's HBMs have all been sold out, and the company has begun to prepare for 2025.