SK hynix announced that it has started mass production of HBM3, which is currently the world’s highest-performance DRAM, and will consolidate its leadership in the high-end DRAM market. Compared with traditional DRAM, HBM is more competitive in terms of data processing speed and performance and is expected to be adopted by more products.
Nvidia has recently completed a performance evaluation of SK Hynix HBM3 samples and is expected to ship to Nvidia systems in the third quarter of this year. According to Nvidia’s schedule, SK Hynix increased HBM3 production in the first half of this year. NVIDIA will apply HBM3 to the H100, which is currently the most powerful accelerated calculator and is expected to improve its accelerated computing performance.
Nvidia purchased HBM3 from SK Hynix with a bandwidth of 819 GB/s, which is in line with the HBM3 high-bandwidth memory standard released by JEDEC earlier this year. According to past information, SK hynix currently offers two capacities. One is 24GB (196Gb) vertically stacked with 12-layer through-silicon via technology, and the other is 16GB (128Gb) stacked with 8 layers, both of which provide 819 GB/s of bandwidth, and the chip height of the 24GB (196Gb) vertically stacked is only 30 microns. Compared with the 460 GB/s bandwidth of the previous generation HBM2E, the bandwidth of HBM3 has increased by 78%. In addition, HBM3 memory also has built-in on-chip error correction technology, which improves the reliability of the product.
SK hynix has been very active in the research and development of HBM. As early as last June, it demonstrated the first HBM3, which provided a bandwidth of 665 GB/s. Then in October last year, SK Hynix announced that it had successfully developed HBM3 memory, becoming the first company in the world to develop a new generation of HBM memory. At ISSCC 2022 earlier this year, SK hynix also
demonstrated HBM3 with a bandwidth of 896 GB/s.
