Samsung, SK Hynix to Increase Share of 8-High HBM4 Shipments, Signaling Changes in Nvidia’s Supply Chain (Zdnet Korea) Samsung Electronics and SK hynix are expected to increase the share of 8-high products in their shipments of sixth-generation high-bandwidth memory (HBM4) to Nvidia. The main driver is Nvidia’s procurement strategy of securing a broader range of HBM options to address thermal-management concerns and improve supply-chain stability. The next-generation HBM4E product is also increasingly likely to be centered on 8-high configurations. According to industry sources on the 26th, Samsung Electronics and SK hynix plan to increase their supply of 8-high HBM4 products to Nvidia in the second half of this year. The two companies had previously been supplying 12-high HBM4 products for Nvidia’s next-generation Vera Rubin series of AI accelerators. HBM is a type of memory in which multiple DRAM dies are vertically stacked, meaning that a 12-high product contains 12 stacked DRAM dies. Samsung Electronics is expected to begin full-scale mass-production shipments in the first quarter of this year, followed by SK hynix in the second quarter. Recently, however, Nvidia’s approach to its HBM supply chain has begun to change. Based on accounts from sources inside and around Samsung Electronics and SK hynix, the two companies plan to gradually increase the share of 8-high products, rather than 12-high products, in their HBM4 shipments to Nvidia during the second half of the year. “Nvidia recently asked suppliers to change the primary HBM4 product configuration from 12-high to 8-high,” said a memory semiconductor industry source. “My understanding is that supply plans for the second half of the year were revised accordingly.” Another source said, “There had already been some demand for 8-high HBM4, but demand for the product increased for the second half of the year, which also affected orders for related materials and components. It is still too early to determine the exact mix, but the trend toward a higher share of 8-high products in HBM4 shipments to Nvidia will become increasingly clear.” The industry believes that a combination of factors may be behind the change. From a performance perspective, thermal issues have emerged as a concern. HBM4 doubles the number of input/output channels—the pathways used for data transmission—to 2,048 compared with the previous generation, substantially increasing heat generation not only at the individual chip level but also across entire server-rack systems. Supply-chain considerations are also believed to have played a role. HBM is currently in extremely tight supply because of aggressive AI-infrastructure investment by global Big Tech companies. At the same time, the improved performance and density of HBM4 make it more difficult to achieve high yields. While yields for the core DRAM dies used in HBM can be improved relatively quickly, yields decline significantly during the processes of stacking 12 dies, bonding them and packaging the finished product. Nvidia is therefore believed to be seeking greater procurement volumes of 8-high products alongside its existing 12-high HBM4 supply, giving it a wider range of memory-capacity options for future AI accelerators. “Nvidia is placing the highest priority on thermal management when designing its latest AI platforms,” said a semiconductor industry source. “This is not so much about downgrading individual chip specifications as it is about finding the optimal system-level configuration, so it is unlikely to cause a major change in overall market demand.” HBM4E, the successor to HBM4, is also increasingly likely to be shipped primarily in 8-high rather than 12-high configurations. HBM4E is expected to be used in Nvidia’s next-generation Rubin Ultra AI accelerator. “Although 12-high and 16-high HBM4E samples are also being discussed, based on the current discussions with customers, the HBM4E products supplied next year are likely to be predominantly 8-high,” said a senior memory semiconductor industry source. “However, uncertainty remains high across the industry because issues such as whether Rubin Ultra will use HBM4E or HBM4 have yet to be resolved.” $SKHY $MU $DRAM

Samsung, SK Hynix to Increase Share of 8-High HBM4 Shipments, Signaling Changes in Nvidia’s Supply Chain (Zdnet Korea) Samsung Electronics and SK hynix are expected to increase the share of 8-high products in their shipments of sixth-generation high-bandwidth memory (HBM4) to Nvidia. The main driver is Nvidia’s procurement strategy of securing a broader range of HBM options to address thermal-management concerns and improve supply-chain stability. The next-generation HBM4E product is also increasingly likely to be centered on 8-high configurations. According to industry sources on the 26th, Samsung Electronics and SK hynix plan to increase their supply of 8-high HBM4 products to Nvidia in the second half of this year. The two companies had previously been supplying 12-high HBM4 products for Nvidia’s next-generation Vera Rubin series of AI accelerators. HBM is a type of memory in which multiple DRAM dies are vertically stacked, meaning that a 12-high product contains 12 stacked DRAM dies. Samsung Electronics is expected to begin full-scale mass-production shipments in the first quarter of this year, followed by SK hynix in the second quarter. Recently, however, Nvidia’s approach to its HBM supply chain has begun to change. Based on accounts from sources inside and around Samsung Electronics and SK hynix, the two companies plan to gradually increase the share of 8-high products, rather than 12-high products, in their HBM4 shipments to Nvidia during the second half of the year. “Nvidia recently asked suppliers to change the primary HBM4 product configuration from 12-high to 8-high,” said a memory semiconductor industry source. “My understanding is that supply plans for the second half of the year were revised accordingly.” Another source said, “There had already been some demand for 8-high HBM4, but demand for the product increased for the second half of the year, which also affected orders for related materials and components. It is still too early to determine the exact mix, but the trend toward a higher share of 8-high products in HBM4 shipments to Nvidia will become increasingly clear.” The industry believes that a combination of factors may be behind the change. From a performance perspective, thermal issues have emerged as a concern. HBM4 doubles the number of input/output channels—the pathways used for data transmission—to 2,048 compared with the previous generation, substantially increasing heat generation not only at the individual chip level but also across entire server-rack systems. Supply-chain considerations are also believed to have played a role. HBM is currently in extremely tight supply because of aggressive AI-infrastructure investment by global Big Tech companies. At the same time, the improved performance and density of HBM4 make it more difficult to achieve high yields. While yields for the core DRAM dies used in HBM can be improved relatively quickly, yields decline significantly during the processes of stacking 12 dies, bonding them and packaging the finished product. Nvidia is therefore believed to be seeking greater procurement volumes of 8-high products alongside its existing 12-high HBM4 supply, giving it a wider range of memory-capacity options for future AI accelerators. “Nvidia is placing the highest priority on thermal management when designing its latest AI platforms,” said a semiconductor industry source. “This is not so much about downgrading individual chip specifications as it is about finding the optimal system-level configuration, so it is unlikely to cause a major change in overall market demand.” HBM4E, the successor to HBM4, is also increasingly likely to be shipped primarily in 8-high rather than 12-high configurations. HBM4E is expected to be used in Nvidia’s next-generation Rubin Ultra AI accelerator. “Although 12-high and 16-high HBM4E samples are also being discussed, based on the current discussions with customers, the HBM4E products supplied next year are likely to be predominantly 8-high,” said a senior memory semiconductor industry source. “However, uncertainty remains high across the industry because issues such as whether Rubin Ultra will use HBM4E or HBM4 have yet to be resolved.” $SKHY $MU $DRAM

Source:https://x.com/jukan05/status/2092484503080206760