Until now, High Bandwidth Memory (HBM) memory was mainly used in AI servers, centers with large databases and the most powerful GPU accelerators due to extremely high data flow, but Samsung is now trying to bring this technology down to the mobile segment as well.
The problem with classic mobile memory is the limited number of I/O channels and large signal losses when the operating speed is increased. Samsung is therefore developing a completely new architecture based on vertical copper conductors and FOWLP packaging technology, which the company already uses in advanced Exynos chips.
The new technology uses a Vertical Copper Post Stack system where memory layers are placed in a kind of “stepped” structure to increase data flow and better thermal efficiency.
HBM memory could drastically speed up AI phones
Samsung allegedly managed to increase the ratio of the height and width of the copper lines from the previous 3-5:1 to even 15-20:1, which significantly increases the data flow in the memory.
However, there is a serious technical challenge as very thin copper lines can bend or crack when the diameter drops below 10 micrometers. That’s why Samsung uses Fan-Out Wafer Level Packaging technology, which additionally strengthens the entire structure and at the same time increases the number of I/O connections.
SamsungReports claim that the new architecture could bring about 30% higher data throughput compared to existing solutions used for the memory subsystem in mobile devices. It is not yet known when the mobile HBM will first appear on the market, but some sources mention the future Exynos 2800 and Exynos 2900 chips as the first candidates.
It is also mentioned that Apple is investigating HBM technology for future iPhone models, while Huawei is reportedly working on similar solutions.
However, the biggest problem currently remains the cost of memory. Due to the huge price increase of DRAM and NAND chips in recent years, phone manufacturers are very cautious about introducing HBM memory into mobile devices.
Samsung is practically trying to turn phones and tablets into much more serious local AI platforms that would perform part of the AI processing directly on the device without relying on cloud infrastructure, reports Wccftech.