Korean semiconductor industry insiders revealed that Samsung's fifth-generation 12-layer high-frequency wide memory HBM3E products finally passed Nvidia quality certification tests and are expected to start supplying high-level memory chips soon, and are expected to enter the next generation of HBM4 competition chain.

The demand for high-efficiency computing market in the global market has increased rapidly, especially catalyzed by generative AI models, supercomputers, and digital twins. The market demand for high-frequency, low-latency memory such as HBM has also increased.

HBM3E DRAM is a key memory technology for solving AI computing power bottles. With a speed of more than 1 TB/s, it has become a necessary component for NVIDIA, AMD, and Intel's new generation AI chips (such as NVIDIA H200/B100/B200, AMD MI350X). Without HBM3E, the computing power of these chips cannot be fully released.

As one of the three major memory manufacturers in the world, Samsung has been relatively behind the HBM competition in the past. SK Hynix was the first to start mass production of HBM3E 12-layer chips in September last year, followed by Micron's certification through NVIDIA in the first quarter of this year. Although Samsung has shipped HBM3E 12-layer chips to Ultramicro, it has not been able to meet NVIDIA's strict performance requirements since it provided NVIDIA's first batch of HBM3E 12-layer chip samples to NVIDIA in February last year.

For this kind of pulling for 18 months, several Korean media recently reported to industry insiders that Samsung has finally officially passed the NVIDIA HBM3E 12-layer chip quality certification test and will become the third company to provide HBM3E 12-layer products to NVIDIA.

However, HBM usually makes reservations one year in advance, while SK Hynix and Micron's all allocations have been sold out this year. Samsung, which is 1 and a half years late, predicts short-term supply volumes to be limited, and the results are to be observed. However, through NVIDIA certification, it mainly represents the end of the long-term card situation and gets the "entry ticket" of the next generation of high-frequency wide memory, which narrows the gap between SK Hynix and Micron's competition in the sixth generation of high-frequency wide memory HBM4, and even some technical indicators may overtake.

HBM4 is expected to debut in NVIDIA's new generation graphics architecture Vera Rubin. NVIDIA requires suppliers to increase the transmission speed of HBM4 data to above 10 Gbps per second. Sources revealed that Samsung has shown 11 Gbps transmission speed, surpassing SK Hynix's 10 Gbps, and Micron has not met NVIDIA's requirements. Samsung plans to ship HBM4 products to NVIDIA this month, hoping to obtain NVIDIA certification as soon as possible.