In a single-chip with 24GB of data.
SK Hynix said that it developed its 12-Hi 24-GB HBM3-6400 memory stacks consisting of 12 DRAM devices connected using via wireless vias (TSVs) and placed on a base dieless than a month ago. IP companies already offer the HBM3 controllers and PHY, and there are a few SoCs that will use HBM3 memory, but none of them are ready to be showcased. So, SK Hynix decided to show the actual 25GB memory chips with a 6.4 GT/s data transfer rate and a 1024-bit interface that can provide up to 819 GB/s. The demonstration indicates that SK Hynix at least has HBM3 samples and can demonstrate them outside the lab. Nevertheless, for obvious reasons SK Hynix can’t disclose the names of its SoC partners that plan on using these memory storages first. Both hardware and software use bandwidth-hungry chips, while other applications use hardware and software like computer computers, so it is definitely better to make sure that most of the first soCs using HBM3 are very expensive.
Samsung’s H-Cube is the case of a smartphone.
But in terms of the costs of HBM, there are many reasons why solutions that require HBM are expensive. First, building high-capacity memory chips is expensive and stacking them all together on a base die makes such known good stacked dies (KGSDs) very precious. Secondly, putting these two blocks together by a substrate is too expensive, since the number of memory stacks that are used increases a lot of time. This is where Samsung and Amkor have introduced a new technology called “Hybrid-Substrate Cube” and promises to make simpler packages with more than six HBM stacks. Interposers are commonly placed on flat-panel substrates (which are generally smaller than interposers, but also very expensive) that are than installed on motherboards. Samsung’s H-Cube Technology applies large and high-density interconnection substrate(HDI) that sits between the Fine-Pitch substrate (FPS) and the motherboard. Since the FPS does not have to be very large enough to be placed directly onto the motherboard, solder ball pitches can be reduced by 35%.Samsung claims its technology has a stable power supply and minimizesthe loss or distortion of multiple logic chips and machines. Samsung stresses that its H-Cube technology works better with higher-end HBM implementations with six or more memory chips. Don’t ever stop anybody from buying more advanced products. Both HBM3 and H-Cube should help improve the bandwidth and adoption of HBM technologies going forward. Were there new products, like AMD’s R9 Fury and RX Vega, going to see them using HBM? A definitely, although that wouldn’t happen for another generation of GPUs.
