ECTC 2026: AI Packaging Hits Quadruple Bottleneck — Power Delivery, HBM4E Routing, Thermal Management & Optical Interconnects
0xBroomberg
Intel, Marvell, and TSMC disclosed four near-commercial solutions at ECTC 2026, the packaging field's top conference. AI accelerator packaging has moved past capacity shortages into four parallel physics problems: power delivery, routing, thermal, and optical.
What does Intel's EMIB-T actually solve?
Intel validated EMIB-T at 36 µm bump pitch — a 65% density increase over the 45 µm pitch used in Granite Rapids. This means → more connection points per unit of bridge silicon, wider data paths between chips.
The "T" stands for through-silicon vias (TSVs — vertical wires drilled through the silicon), and the core purpose is power delivery. Intel says bridge TSVs cut DC voltage drop by 68-80%. In plain terms = far less power is lost between the supply and the chip, so the die sees a steadier voltage.
The bridge also adds MIM capacitors — metal-insulator-metal energy-storage structures — at 500 nF/mm², improving AC impedance of the power distribution network by more than 82%. This means → the voltage spikes that occur during rapid switching are heavily smoothed.
Intel is also testing 25 µm pitch, but below 25 µm solder volumes become tiny, raising short-circuit and yield-loss risks sharply — the limiting factor shifts from routing density to manufacturing precision.
How is the HBM4E routing problem being tackled?
HBM4E — the enhanced version of fourth-generation high-bandwidth memory, faster but harder to route — has signal paths of varying lengths. The longest channels are most vulnerable to crosstalk and insertion loss. Intel's answer: route the longest signals on the cleanest wiring layers.
On the M9 layer, only about 28% of the longest channel passes through the tightest routing zone; on lower layers like M3, that share rises to 84%, but those channels are physically shorter, limiting the damage. In plain terms = long wires get the wide road, short wires take the narrow one, so the worst path does not drag down the whole link.
Signal-integrity simulations show eye width of about 67% UI at 12 Gb/s, rising to roughly 72.5% with a one-tap DFE equalizer. Even at 16 Gb/s, eye width stays above 60%. This reflects meaningful signal margin at high speeds.
What does Marvell's "custom HBM" change?
Standard HBM interfaces are defined by the JEDEC spec — good for interoperability, but constraining for power, performance, and area. Marvell's approach: keep the DRAM core die unchanged, replace the standard base die with a custom one built on an advanced logic process, integrating the HBM controller, monitoring, and an expanded interface.
This means → the PHY (physical-layer interface circuits) and related logic that the host ASIC normally dedicates to HBM shrink by roughly 60% — freeing that silicon area for more compute.
At the package level, the custom interface shortens interposer channel length from 6.5 mm to 1.5 mm. With the same 9-layer routing and 2/2 µm line width/spacing, bandwidth climbs. Marvell's example uses 1,024 lanes at 32 Gb/s, reaching 4.1 TB/s.
Nvidia has announced that its next-generation Feynman will use custom high-bandwidth memory. Analysis estimates that roughly 16% of the current Rubin GPU die area is dedicated to HBM-related logic — the custom approach can offload much of that burden onto the HBM base die.
Why has cooling become a hard wall?
Under conventional cooling at 1-2 liters per minute (LPM), lidded packages dissipate 1.9-2.3 kW and lidless packages 2.5-3.0 kW, using 40 °C deionized water. In plain terms = the ceiling for today's cooling is roughly 3 kW.
Both configurations saturate above 4 LPM — more flow does not remove more heat. This means → for next-generation multi-kilowatt AI chips, conventional cooling is no longer sufficient.
TSMC and Microsoft are exploring microfluidic cooling: routing coolant directly into the silicon itself, removing heat at the source rather than waiting for it to conduct to the package surface.
What can optical interconnects solve?
Marvell demonstrated an optical multi-chip interconnect bridge (OMIB) that replaces electrical signals with light for die-to-die links inside the package. This means → the link is no longer confined to the package edge; a single bridge structure handles both intra-package die-to-die and external optical interconnects.
Lightmatter also presented an in-package optical interconnect solution at ECTC, indicating multiple vendors are advancing this path in parallel.
Samsung proposed an 8-layer silicon interposer that it says reduces layer count by 20% versus estimated requirements, with 75% of layers dedicated to signal routing in an alternating dual-signal / single-ground shielding pattern.
How close are these solutions to volume production?
Intel's EMIB-T is scaling from 2× reticle silicon to 4.5× reticle package validation, targeting certification by end of 2026. But demo samples showed visible warpage; substrate handling and overlay accuracy remain first-order constraints.
Intel acknowledges EMIB-T trails TSMC's CoWoS platform on multiple dimensions — TSMC has already deployed deep-trench capacitor integration and is further ahead on integrated voltage regulators and active local silicon interconnects.
Analysts note that achieving low-warpage, crack-free hybrid bonding — a process that joins two chips directly face-to-face — requires coordinated optimization of copper, dielectrics, CMP, surface preparation, and anneal. The pace of material and equipment supplier improvements from 2027 onward will be the key checkpoint for whether AI packaging bottlenecks can be systematically broken.
Content is for reference only, not financial advice.