@zephyr_z9: Well tHBM has a huge power delivery problem U will have to route 1000s of Amps thru the HBM stack
Summary
A discussion highlights the significant power delivery problem in tHBM technology, requiring routing of thousands of Amps, and critiques Samsung's zHBM for potential melting issues due to stacking memory on hot GPUs, as revealed by HBM creator Kim Jung-ho.
View Cached Full Text
Cached at: 08/19/26, 04:37 AM
Well tHBM has a huge power delivery problem U will have to route 1000s of Amps thru the HBM stack
Fireside Alpha (@firesidealpha): HBM creator Kim Jung-ho reveals his lab’s tHBM flipped stack publicly for the first time saying Samsung’s zHBM with memory stacked on a hot GPU simply melts and is likely not the right fundamental direction
“So the idea is, let’s just put the HBM on top of the GPU. Then you
Similar Articles
Hot Chips 2026: Samsung and HBM Base Die Opportunities
At Hot Chips 2026, Samsung presented opportunities to utilize unused area on HBM base dies after moving to a logic node, with phases including integrating memory controllers to optimize PHY area and power.
Micron: HBM Requires Three Times More Wafer Area Than DDR5
Micron states that High Bandwidth Memory (HBM) requires three times more wafer area compared to DDR5 memory.
What Comes After HBM (7 minute read)
The memory industry is experiencing a supercycle due to high AI demand for HBM in Nvidia GPUs, with supply shortages expected to continue and memory manufacturers projected to achieve high profits despite cyclicality.
@tobi: worth noting that the memory bandwidth of a GB300 is 7.4 TB/s. That's why HBM memory is hard to get on planet earth rig…
The tweet highlights the high memory bandwidth of GB300 (7.4 TB/s) and explains that this contributes to the scarcity of HBM memory globally.
High-Bandwidth Flash offers efficient storage for model weights
IEEE Spectrum reports on High Bandwidth Flash (HBF), a new memory technology that stacks NAND flash dies using high-bandwidth memory (HBM) packaging techniques to deliver vastly higher read bandwidth for efficient storage of LLM model weights, with first products expected in about a year.