NVIDIA seems to have decided to step into the balance of power in the HBM memory market, which is currently dominated by Samsung, SK Hynix and Micron. The company is betting that its NVHBM architecture can significantly change the balance of influence for manufacturers.
The main difference with NVHBM is that the memory controller is moved directly from the main XPU die to the base layer of the HBM stack. According to reports, this can increase throughput by approximately 30% while reducing power consumption by 15% relative to HBM4E. In addition, up to 25% of the main chip area is freed up for computing units.
In this context, reports that SK Hynix plans to modify basic crystal production are particularly interesting. The company now produces them using TSMC’s 12-nanometer process technology, which costs three to four times the cost of traditional DRAM. According to Korean media reports, SK Hynix is considering moving production to Intel to reduce costs.
For NVIDIA, this isn’t just an attempt to make memory faster and more affordable. The controller became one of HBM’s most valuable parts, so the company was essentially trying to control the entire design. If the idea works, DRAM manufacturers risk switching from being major owners of the market to component suppliers, while NVIDIA would gain more control over HBM costs and development.
