HBM scaling runs into limits in thinning, TSV density and thermal control

August 27, 2026

HBM scaling runs into limits in thinning, TSV density and thermal control

Image: ChatGPT

HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. scaling is no longer mainly a matter of adding capacity. At Hot Chips 2026, speakers from Micron, Samsung and SK hynix described how each additional DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology layer pushes the technology into a harder set of tradeoffs involving thinner dies, more TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures. area, thermal dissipation and limited fab capacity.

To stay within HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. package height limits, the dies in a stack must be thinned further, but that makes handling, wafer alignment and bonding more difficult. Thinner dies are also more prone to warpage. At the same time, the number of TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures.s keeps rising because they are the main data paths in 3D-IC3D Integrated Circuit3D-IC refers to integrated circuits with three-dimensionally stacked components. HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. modules and are central to bandwidth scaling. The problem is that TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures.s consume valuable die area, reducing the space available for memory cells on each layer.

Samsung said HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. bandwidth has roughly doubled from generation to generation, but the bottlenecks are now the TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures. network between dies and the PHY I/O in the base die. Increasing TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures. count is straightforward compared with speeding up the TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures.s themselves, yet the growing TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures. array forces continued pitch reduction and consumes more silicon area. Micron added that HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth.’s peripheral circuitry and TSV1. through silicon viaVertical conductive connection through silicon used to interconnect stacked dies.2. through-silicon viaAn interconnect technology used in HBM and HBF structures.s are what distinguish it from conventional DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology, even though the memory cell technology is still commoditized DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology.

The manufacturing burden is showing up in supply. Because HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. uses far more silicon and wafer capacity per bit than standard DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology, it is putting pressure on the same production lines that make conventional DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology. Industry speakers said yield is being hit by underfill voids, misalignment and process variation, any of which can turn an entire stack into expensive scrap. The result is a much higher appetite for wafers at a time when new DRAM1. dynamic random access memoryMemory technology or system for temporarily or permanently storing digital information.2. dynamic random-access memoryDRAM is a volatile memory technology fabs still take years to build.

Thermals are becoming a separate design constraint. In a stacked HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. cube, the hottest logic die sits near the base, while cooling is applied at the top, so heat has to travel through the full stack. SK hynix said its 16-high work exposed additional issues around cube thickness, gap height, die-wall pitch and oxide layers. Micron said the industry is already looking beyond today’s thermal-compression-bonded microbumps toward fusion bonding and hybrid bonding, which could support tighter pitches and improve thermal resistance.

Micron also pointed to future high-speed I/O changes, including memory-optimized SerDes1. Serializer/DeserializerSerDes is the high-speed I/O block that receives, conditions, and retransmits the signal.2. Serializer-DeserializerSerDes converts parallel signals into serial data streams and back. PHYs and, eventually, optical links. For now, HBM1. human body modelESD test model that simulates a person discharging static electricity into a component.2. High Bandwidth MemoryHBM is a stacked DRAM architecture designed to provide very high memory bandwidth. scaling remains constrained by the combination of density, bandwidth and package physics, with more layers creating more ways for the stack to fail.

Source: Semiconductor Engineering

More background: How does it work?

← All posts