It has been confirmed that Samsung Electronics plans to begin shipping 5th generation high-bandwidth memory (HBM3E) products to Broadcom as early as the second half of this year. However, the delivery of the 12-layer HBM3E product, which 'big player' Nvidia is currently conducting quality tests on, has been delayed. While SK hynix and Micron are supplying HBM to Nvidia ahead of Samsung Electronics, the schedule for Samsung Electronics, which aimed to supply as early as the second quarter, is being delayed more than expected.
There are three main technological factors hindering Samsung Electronics' delivery to Nvidia. The first is that it has not met Nvidia's thermal standards, which are about twice as high compared to Broadcom's. The second is the possibility of data transmission issues arising when HBM is operated under extreme conditions, such as overheating, raising concerns that errors could occur during AI computation processes in Nvidia's high-speed data communication system, 'NV Link.' The third is assessed to be Samsung Electronics' low yield (the ratio of good products to production volume) compared to its competitors.
An industry insider noted, "It is true that the supply chain entry was delayed compared to SK hynix and Micron due to three main issues. However, it has recently been reported that most of the technical challenges have been overcome. Nevertheless, since Nvidia has already allocated supply volumes to SK hynix and Micron, negotiations for supply with Samsung Electronics appear to be somewhat delayed."
◇ Nvidia's thermal standards are stricter than Broadcom's
According to industry sources on the 18th, Samsung Electronics plans to supply HBM3E to the application-specific integrated circuit (ASIC) design company Broadcom starting in the second half of the year. Samsung Electronics is expected to account for more than 50% of Broadcom's HBM3E supply chain, effectively becoming the largest supplier of HBM3E to Broadcom.
It is understood that the thermal standards required by Broadcom are less than half of Nvidia's standards. Unlike Nvidia's chips, AI semiconductors designed by Broadcom are customized to optimize each company's needs. They are designed and manufactured in accordance with the performance requested by global tech giants like Google and Meta for final delivery.
However, Nvidia's AI semiconductors are produced as universal products that can be utilized across various applications. This means they must demonstrate high performance in any environment, which leads to high power consumption. Excessive power usage can exacerbate the thermal phenomena in chips. If the HBM mounted on them generates excessive heat, the overall performance of the chips is bound to decline. Since Nvidia's AI semiconductors boast industry-leading performance, their thermal standards are also reported to be high compared to competitors like AMD. It has been communicated that Samsung Electronics has yet to meet Nvidia's thermal standards.
Some point out that when the Samsung Electronics HBM is mounted, data transmission speed and accuracy in extreme conditions, such as overheating, lag behind those of SK hynix and Micron. The key to the competitiveness of AI semiconductor performance is to support AI in efficiently learning large volumes of data. For this, it is necessary to send large amounts of data as quickly as possible without bottlenecks, enabling the Nvidia graphics processing unit (GPU)—which serves as the brain of the AI semiconductor—to learn this efficiently. In particular, since Nvidia employs a high-performance proprietary system called NV Link, concerns have been raised that there may be problems with data transmission in this case.
Nvidia introduced the NV Link system to ensure that each chip responds faster in server environments where AI semiconductors are heavily deployed, without experiencing data bottlenecks in any environment. However, due to performance issues with DRAM mounted on Samsung Electronics HBM, some point out that the probability of facing data transmission speed or accuracy issues could be higher compared to SK hynix and Micron. If the processor cannot accurately recognize the data transmitted through HBM, it will impact the performance of AI models powered by the AI semiconductors.
◇ Samsung Electronics is launching a full-scale campaign to improve quality through DRAM redesign and yield stabilization
The low yield of Samsung Electronics' HBM has been identified as a problem. When yield is low, it becomes difficult to supply the contracted volume on time, and competitiveness diminishes during price negotiations. Accordingly, Samsung Electronics is known to be working hard to redesign DRAM to improve quality and stabilize yield compared to the past. Additionally, to enter the Nvidia supply chain, it is reported that the prices of previously produced stock volumes have been set lower than those of competitors.
An insider in the semiconductor industry said, "The yield issue, which has been pointed out alongside the thermal problem for a long time, has been rapidly improved. Samsung is in negotiations to enter the Nvidia supply chain and are also accelerating delivery by sending next-generation HBM samples."