News

South Korea and Japan take action at the same time! DRAM chips placed vertically: AI memory heat dissipation problem solved

2 min read
Kuai Technology reported on July 11 that Korean and Japanese researchers simultaneously proposed two new HBM (high-bandwidth memory) solutions at the 2026 IEEE/JSAP VLSI Technology and Circuit Symposium. The core ideas are surprisingly consistent: put DRAM (dynamic random access memory) chips upright to solve the increasingly serious heat dissipation bottleneck of AI chips. The Korean plan is called V-Die and was proposed by the Ulsan National Institute of Science and Technology (UNIST). This design rotates the custom DRAM chip 90 degrees and places it vertically, using TSV (Through Silicon Via) to free up the chip area to accommodate more memory units. The bottom of each chip has an independent I/O channel, and liquid cooling channels are plugged between adjacent chips. In the simulation test, under the same capacity, V-Die ran GPT-3 level load and reached 540 tokens/s, while traditional HBM4 only had 296 tokens/s, nearly doubling the performance. The Japanese solution is called MOSAIC, led by a team from the University of Tokyo. The idea is similar but the focus is different. It uses orthogonal chip stacking and contactless inter-chip interfaces. Data transmission does not rely on physical contacts, but is coupled through micro-inductive coils. The rate of each channel reaches 4 Gbps. The research team claims that this structure can achieve twice the capacity of HBM4 in a DRAM-on-GPU configuration. The two solutions address the same pain point. The AI ​​chip is stuck in memory. Modern AI accelerators have sufficient computing power, but large models must frequently move massive amounts of data between memory and computing units. HBM alleviates this contradiction by vertically stacking DRAM chips close to the processor. NVIDIA Blackwell Ultra B300 is equipped with a 288GB HBM3E. But the higher they are stacked, the more difficult it is to dissipate heat. V-Die and MOSAIC essentially use "side placement" to exchange heat dissipation space, and by the way, the bandwidth and capacity are also increased. However, these two technologies are still in the paper stage and are still far away from mass production.