You have to catch up with the leader.
According to analysts surveyed by Reuters, this year SK hynix will control more than 80% of the supply of HBM3 and higher memory for NVIDIA's needs, and this is largely due to the ability of this South Korean manufacturer to produce memory chips with better characteristics than competitors. The secret, according to the source, lies in the use of mass reflow with partial mold filling (MR-MUF) technology when forming insulation between memory layers in HBM chips. Among other things, this technology makes it possible to increase the efficiency of thermal energy removal from the memory chip and reduce the operating temperature by more than 10 degrees Celsius, which is critical for the operating conditions of computing accelerators, which are equipped with such memory.
Image source: SK hynix
Samsung Electronics, for example, has so far used thermal compression non-conductive film technology to create such insulation between memory layers in a stack, but it is inferior in properties to the MUF technology used by SK hynix since 2021. Secondly, the latter company manages to obtain up to 70% of usable chips during the production of HBM3, while for Samsung this figure does not exceed 20%.
Now, as reported by Reuters, Samsung has decided to follow the example of SK hynix and begin introducing a new method of packaging HBM3 chips using MUF technology. For this, the necessary equipment and consumables are purchased, but Samsung will be able to supply memory chips created using MUF in mass quantities only next year. Until then, the company will combine both technologies on its production line. Samsung is making similar efforts to receive orders from NVIDIA for the supply of HBM3 and HBM3E memory to equip current generations of computing accelerators. This year, the capacity of the HBM memory market will more than double to $9 billion amid the boom in artificial intelligence systems, according to TrendForce analysts.