Samsung is expected to more than double output of its HBM4 and HBM4E DRAM
summary
Samsung Electronics is planning to significantly increase the production of its HBM4 and HBM4E high-bandwidth memory chips next year to meet growing AI demand. The company aims to more than double its output of these specific chips by increasing the use of essential glass carriers used in the wafer thinning process. This strategic shift focuses on high-value, high-layer stacking products to maintain competitiveness in the AI accelerator market.
key points
- Samsung will increase its outsourced cleaning volume for glass carriers from 20,000 to 50,000 sheets per month, indicating a massive ramp-up in production capacity.
- The production mix is expected to shift heavily toward HBM4 and HBM4E, which will likely make up 80% of total HBM output next year compared to 40% this year.
- The company is focusing on 12-layer and higher stacks, leveraging 10-nanometer-class DRAM and 4-nanometer base dies to enhance memory performance for AI customers like Nvidia.
who should read it
Investors and semiconductor industry analysts who need to track Samsung's production capacity and technological trajectory in the AI chip race.
discussion positive
consensus
Commenters agree that High Bandwidth Memory (HBM) production, rather than logic chips or EUV equipment, is the primary bottleneck for China's AI accelerator capabilities. While sanctions force China to develop its own HBM ecosystem, low yields and complex packaging requirements currently hinder progress. There is a general sentiment that these restrictions may ultimately accelerate China's self-sufficiency by creating necessity.
top insight
The discussion clarifies that HBM is unsuitable for general consumer CPUs because it requires a massive 2048-bit bus and prioritizes bandwidth over latency. Since standard CPUs operate on small cache lines and are highly sensitive to latency, the architectural shift to HBM would require a complete motherboard and processor redesign that offers little benefit for non-GPU workloads.
Generated by an LLM from the article and thread. Check the sources before relying on it.