In the past year, news about domestic companies' entry into the memory industry has been buzzing. With the original Infineon, Ziguang has already made breakthroughs in DDR3 memory, producing DDR3 memory in small batches and introducing more mainstream in the second half of the year. DDR4 memory chips are trying to catch up with the international mainstream.
However, looking at the entire memory market, DDR5 memory will soon be coming. Even more frightening is that even DDR5 memory is likely to be replaced by newer technologies. The industry has already proposed that DDR memory will die. In the future, high bandwidth is needed. The product will switch to HBM memory, with HBM 3 memory in 2020, HBM 4 memory in 2024, bandwidth up to 8TB/s, and single-slot capacity up to 512GB.
For HBM memory, DIY players can be quite familiar with it. AMD first commercialized the first generation of HBM technology on the 2015 Fury graphics card. The ultra-high bandwidth, ultra-low area occupation completely changed the graphics design at that time, and then NVIDIA HBM 2 technology was commercialized on the Tesla P100, but the HBM 2 technology used in the consumer market was AMD's RX Vega card last year. However, because the cost of HBM 2 memory is too high, the RX Vega actually uses two sets of 4GB HBM 2 The equivalent bit width is reduced by half compared to the first generation. Although the frequency is greatly increased, the actual bandwidth is lower.
The players who have seen HBM are surely impressed by the technology. How should it develop in the future? Nicolas Dube of HPE (HP Enterprise) shares some of his opinions a few days ago. In his opinion, DDR memory is coming to an end. (DDR is Over), especially in some occasions with high bandwidth requirements.
According to some data he shared, HBM 2 memory will be used in large quantities in 2018, HBM 3 will be applied around 2020, improved HBM 3+ technology will be applied in 2022, and HBM 4 memory will be available in 2024, bandwidth and capacity. It will also increase step by step, such as the current HBM 2 memory, with a core capacity of 8Gb. Through TSV technology, each CPU can support 64GB of HBM2 memory, and the bandwidth of each slot can reach 2TB/s. In the era of HBM 4, each CPU supports a capacity of 512GB and a bandwidth of more than 8TB/s. At present, AMD's EPYC processor supports 8-channel DDR4 memory, although the maximum capacity can reach 2TB, but the bandwidth is about 150GB/s, which is far worse than HBM memory.
According to his point of view, in some occasions requiring high bandwidth, HBM technology will undoubtedly outperform DDR memory, so he said that DDR memory will be dead in this respect. For example, the HPC high-performance computer industry is in great need of HBM. Come back, DDR will die this judgment is not suitable for the desktop market, HBM technology, although all kinds of good, but now look at the cost problem will not be able to solve.
Currently, only Samsung, SK Hynix, and Hyunix can produce HBM memory. Because HMC has HMC technology, it is not very enthusiastic about HBM. Therefore, HBM's cost reduction process will be long. For desktop players, DDR4 is a long time Not going out of date, DDR5 memory will start to be implemented around 2020, so we can't see DDR memory being destroyed by HBM within three to five years.
