HBM War Round 2: Samsung zHBM vs SK Hynix (2026)
HBM (High Bandwidth Memory) is stacked DRAM that feeds data to AI processors at extreme speed. Because AI performance now stalls on memory rather than raw compute, whoever builds the best HBM sets the pace for the entire AI industry. In presentations revealed on August 24, Samsung and SK Hynix collided head-on at Hot Chips 2026.
Samsung's zHBM: Stacking Memory Directly on the GPU
Samsung laid out a three-phase roadmap that ends in zHBM. The design removes the interposer (the base plate connecting chips) and stacks DRAM vertically on the processor itself. Samsung claims the architecture delivers 70% better power efficiency and 230% more bandwidth than a standard HBM4E stack, while its Heat Path Block cuts peak temperature by more than 35%.
Why SK Hynix Pushed Hybrid Bonding to HBM5
SK Hynix chose patience. Hybrid bonding, which joins copper wiring directly to shrink the gap between chips, will not arrive until HBM5, with full-scale production expected around 2029-2030. Until then, the company extends its proven MR-MUF process through Nvidia's next-gen Rubin platform and leans on I-HBM cooling, which cuts thermal resistance by more than 30%. Per Tom's Hardware, SK Hynix already holds about 70% of Rubin's HBM orders.
The 0.775mm Ceiling: Where the Real Fight Begins
The most striking number on stage was 0.775mm. A memory stack cannot grow much beyond standard wafer thickness, or 775 microns. Products fitting 16 DRAM layers and 48GB inside that limit are already in customer qualification. With thermal load rising 2.2x per generation, the contest has shifted from stacking higher to bonding and cooling smarter.
Who Holds the Power in the AI Supply Chain?
Analysts read this clash as a rule change for the memory industry. A commodity product is becoming custom silicon, and suppliers are turning into design partners with genuine pricing power. SK Hynix weighing Intel's EMIB packaging alongside TSMC's approach points the same way: alliances across the chip supply chain are in play. For Nvidia, AMD, and US hyperscalers building AI capacity, the leverage now sits with the two memory giants.
Key Takeaways
① Samsung's zHBM - stacks DRAM directly on the GPU, targeting 70% better power efficiency and 230% more bandwidth.
② SK Hynix's defense - hybrid bonding waits for HBM5; I-HBM cooling and 70% of Rubin orders hold the line.
③ The real battleground - taming heat inside a 0.775mm ceiling decides the next crown.
AI's competitive edge is decided as much by hidden components as by visible services. A war fought inside less than a millimeter of silicon is redrawing the map of the global AI industry.
👉 Samsung Foundry Price Hike: 15% Increase as TSMC Runs Full - also worth a read.
📌 Sources: Tom's Hardware, TrendForce (2026)
댓글
댓글 쓰기