SK Hynix Unveils 48GB 16Hi HBM3E, Develops PCIe 6.0 SSD and UFS 5.0

At the SK AI Summit 2024, SK Hynix has announced the release of the world's first 16-high (16Hi) HBM3E memory, leading competitors Samsung and Micron. The company emphasized the memory's "technical stability" and confirmed that samples are expected to be available as early as the beginning of next year.

Just weeks ago, SK Hynix launched its 12-high (12Hi) HBM3E memory, securing contracts with AMD for the MI325X and NVIDIA for Blackwell Ultra. The new 16-high HBM3E memory offers a significant increase in capacity, reaching 48GB (3GB per chip). This higher density allows AI accelerators to achieve up to 384GB of HBM3E capacity in an 8-stack configuration.

SK Hynix claims that the 16-high HBM3E memory enhances AI training performance by 18% and inference performance by 32%. Like the 12Hi version, the 16-high HBM3E uses MR-MUF packaging technology, which connects chips by melting the solder between them.

The company expects to have 16-high HBM3E samples ready by early 2025. However, this memory may be a transitional product, as SK Hynix plans to launch its first 12-high HBM4 products in the second half of 2025, with 16-high HBM4 to follow in 2026.

In a strategic move, SK Hynix also revealed that NVIDIA's CEO Jensen Huang has requested a six-month advance supply of HBM4. SK Hynix has expressed its willingness to meet this request, strengthening its partnership with NVIDIA. This could pave the way for early mass production of HBM4.

Reports suggest that SK Hynix completed the tape-out of HBM4 in October 2024. Following the standard chip development timeline, NVIDIA and AMD are expected to receive certified samples by the first or second quarter of 2025.

HBM4 will feature an increased channel width from 1024 bits to 2048 bits, supporting over 16 vertical DRAM chip stacks with 4GB of memory per chip. This significant upgrade is expected to meet the demands of next-generation AI GPUs.

In addition, SK Hynix is actively developing PCIe 6.0 SSDs, high-capacity QLC (Quad-Level Cell) eSSDs for AI servers, and UFS 5.0 for mobile devices. The company is also working on an LPCAMM2 module and developing soldered LPDDR5/6 memory using its 1c nm process node.

To address the "memory wall," SK Hynix is exploring solutions such as Near-Memory Processing (PNM), Processing-In-Memory (PIM), and Compute-Storage, following in the footsteps of Samsung, which has demonstrated its own PIM version. These innovations aim to process data directly in memory, eliminating the need to transfer data to external processors.