SK Hynix expects that HBM3 will have a bandwidth of 665 GB/s

High Bandwidth Memory, also known as HBM DRAM, is a high-performance DRAM based on 3D stack technology, suitable for application scenarios with high memory bandwidth requirements. Although HBM has not become the mainstream DRAM for graphics cards, it is still seen from time to time in fields such as data centers that require high bandwidth. Recently, SK Hynix announced the HBM3 product plan, introducing information such as the upcoming specifications and expected bandwidth.

Image: SK Hynix

At present, JEDEC, which formulated the HBM standard, has not officially released the HBM3 standard specification. As the main promoter of HBM, SK Hynix has been committed to the research and development of the next generation of HBM. According to a report by TomsHardware, on SK Hynix’s website, HBM3 products show that they are “under development” and have reached an I/O speed of 5.2 Gbps, provides a bandwidth of 665 GB/s. These two values ​​are higher than the 3.6 Gbps and 460 GB/s of HBM2E, an increase of 44.4% and 44.6% respectively. Of course, this is not the ultimate goal of SK Hynix for HBM3. Other manufacturers have announced that they will expand to 7.2 Gbps I/O speed.

Now supercomputing or FPGA-like devices will use 4-6 HBM2E stacks to obtain 1.84-2.76 TB/s bandwidth. If you change to HBM3, you will get a bandwidth of 2.66-3.99 TB/s. In early 2020, SK Hynix obtained the license of DBI Ultra 2.5D/3D interconnection technology from Xperi Corp. It is specially used for high-bandwidth memory solutions. It supports 2.5D and 3D integrated packaging. It can also integrate IP modules of different sizes and different processes and can be used for highly integrated CPU, GPU, ASIC, FPGA, and SoC.

SK Hynix did not disclose the release date of HBM3. It is not expected to enter the consumer market so soon and apply it to graphics cards. In the past, AMD used HBM as video memory on Radeon graphics cards, but currently only plans to use it on acceleration cards with CDNA architecture. For the time being, Intel and NVIDIA’s plans are similar to AMD.