AMD reckons 3D DRAM could cut memory power, but old packaging headaches keep the idea stuck in the lab.
According to Tom’s Hardware Italy, AMD is looking at 3D DRAM combined with hybrid bonding that could be up to 17 times more energy-efficient than conventional HBM stacks using microbumps.
For those not in the know, 3D DRAM involves stacking DRAM and linking it with copper-to-copper connections, rather than relying on tiny solder balls between layers.
Those solder balls, known as microbumps, are used in today’s HBM stacks. They take up space, add electrical resistance and leave gaps that need filling, which is not ideal when AI chips are already guzzling power and throwing out heat.
Hybrid bonding eliminates the bumps. It flattens the surfaces, lines up the copper pads, and bonds the layers directly, reducing the distance signals travel and reducing wasted energy.
According to Semiconductor Engineering, HBM4 is sticking with microbumps for now because hybrid bonding still has cost, yield, testing and surface-cleanliness problems. It said hybrid bonding can offer much lower energy per bit, but getting it into volume production is another matter.
AMD has some form here through 3D V-Cache, where cache memory is stacked on top of a processor instead of sitting beside the compute cores. Moving that sort of thinking into DRAM is trickier.
DRAM cells must hold charge reliably, and the heat used in bonding can degrade refresh behaviour. A bit of contamination only a few nanometres wide can wreck copper-to-copper links, which means production needs brutally clean processes and tight control.
The wider AI memory market is already looking for ways around HBM bottlenecks. SRAM-only decoding, processor-in-memory within LPDDR, CXL pooling, and 3D DRAM are all being explored as vendors try to move more data with less power.
Semiconductor Engineering said commercial 3D DRAM could bring density and energy-efficiency gains for datacentres and AI applications. Thermal management and die-stacking problems still need sorting before anyone can declare HBM yesterday’s problem.
3D DRAM looks like a promising route rather than a shipping product. AMD’s 17 times figure shows why the industry is interested, while the manufacturing mess explains why HBM is not going away just yet.
