"The problem is compounded by the Trump administration's recent ban on Chinese factories manufacturing these components. The largest component manufacturer, Samsung, announced last week in the Korean press that it was going to focus on the production of DRAM chips to the detriment of NAND at its Pyeongtaek and Hwaseong factories in South Korea, which until now have manufactured both types of component."
"The pressure on DRAM manufacturing will have an impact on internal HBM memories in the GPUs designed for AI training. These are the same circuits in both cases - they are etched on the same chains. The difference lies in the way they are assembled on a communication bus. The HBM bus is very large, because it is designed to interface directly with the pins of a computing chip,"
Global demand for server memory from cloud providers is rapidly outstripping supply, driven by large AI model sizes that require tens of gigabytes for inference and training. Samsung has shifted production capacity at its Pyeongtaek and Hwaseong plants toward DRAM at the expense of NAND to meet higher RAM demand. The same fabrication chains produce HBM used inside GPUs, so DRAM pressure threatens HBM availability for AI training accelerators. DDR5 wholesale prices have surged about 60% this month, with expectations of further increases. A US ban on Chinese manufacturing adds further supply constraints.
Read at ComputerWeekly.com
Unable to calculate read time
Collection
[
|
...
]