HBM (High Bandwidth Memory) memories are specially designed for memory-intensive computing tasks. Its main characteristic is bandwidth, much higher than GDDR memories integrated by gaming graphics. Thus, HBM memories offer excellent performance for high-level computing or artificial intelligence.
Precisely, where the demand for these memories is the greatest, it is in specific solutions for artificial intelligence. An example would be the new NVIDIA Blackwell, a GPU architecture directly focused on AI. These solutions integrate large amounts of HMB memory, since they must support large volumes of data.
Serious problems with HMB3 memories
According to various industrial sources, NVIDIA has failed to validate the HBM3 chips manufactured by Samsung. Apparently, the South Korean company would have big problems in the manufacturing of these memories.
Different industrial sources close to the subject highlight a set of shortcomings. The first and most important would be that Samsung’s HBM3 memories are getting hotter than they should. Note that NVIDIA has quite a few problems cooling its most powerful solutions, which only adds to the problems (and temperature, of course).
The most serious problem would concern the power consumption of Samsung’s memory chips. It seems that these HBM3 memories have higher consumption than expected, which creates another problem for NVIDIA.
It appears that Samsung tried to convince NVIDIA to validate the HBM3 and HBM3E memories since at least the end of 2023. Sources suggest the two companies have been working for at least six months to identify the problem and resolve it.
More specifically, the problem would concern Samsung’s 8- and 12-layer HBM3E memories. Due to these issues, NVIDIA is currently reportedly using HBM3 memory chips from Micron and SK Hynix starting in mid-2022 and HBM3E chips from both companies starting in March this year.
What is not made clear is whether this is an issue with Samsung’s DRAM Fabs production plants, an issue with the encapsulation process, or some other unknown issue. According to some industry sources, this could be because Samsung had less time than its competitors to develop HBM memories.
Samsung, in view of this information, published a press release to try to clarify things. The company indicates “the process of optimizing its products through close collaboration with customers” without giving further details.
In addition, the South Korean company emphasizes that “allegations of outages due to heat and electricity consumption are not true.” Note that the company does not provide specific data on this, it only states that the information is incorrect.
At this time, NVIDIA, the other party in this case, has not commented on this matter. It’s possible they won’t, unless Samsung pushes hard to try to prevent damage to the company’s image.
