CoWoS Warpage Explained: Why NVIDIA Blackwell Needs Liquid Cooling (It's the Temperature Swing, Not the Heat)

Bottom line: "Blackwell is liquid-cooled because it runs hot" is only half the story. The harder problem is mechanical. A CoWoS package stacks materials that expand at very different rates, so every idle-to-full-load cycle stretches and relaxes the bond layers. Liquid cooling lowers the temperature and, just as important, shrinks the swing, and that is what buys solder joints and microbumps their lifetime.

What happened

• Power: each GPU on the HGX B200 board is configurable up to 1 kW (NVIDIA docs). In GB200 NVL72 the B200 reaches about 1,200 W and in GB300 NVL72 the B300 about 1,400 W, and there is no air-cooled NVL72 (Barrack AI summary).
• During the 2024 Blackwell delay, SemiAnalysis reported that CTE mismatch between the GPU dies, silicon bridges, organic interposer and substrate in CoWoS-L caused warpage, and that NVIDIA reworked top metal layers and bumps (SemiAnalysis, 3D InCites).
• At OCP APAC in October 2026, TSMC advanced packaging head Jun He described "mismatched gradients" building stress as materials sit closer together, said thermal and warpage difficulty rises nonlinearly past 5.5x reticle, and cited parts that passed component-level tests but showed chipping and low-k delamination in customer systems (Simple Tech Trend recap).

Point 1: Heat flux has outrun air

B200 is two near-reticle-limit dies. Assuming about 1,600 mm² of combined die area, 1,200 W averages roughly 0.75 W/mm², or 75 W/cm², before hotspots (author estimate). That heat has to cross the die, the thermal interface material, the lid and the heatsink, each adding resistance. Air has poor heat capacity and conductivity, so it can only scale with bigger fins and faster fans. A cold plate puts water right against the lid and runs the silicon far cooler at the same power.

Point 2: The real killer is CTE mismatch

Silicon sits near 2.6 ppm/K and organic substrate near 15 ppm/K, a gap of about 12 ppm/K. A 5.5x-reticle footprint is about 4,720 mm², a ~69 mm square whose corners are ~49 mm from center. Rough free-expansion math: a 55 K swing gives ~32 µm of relative movement at the corner; at 35 K it is ~20 µm (author estimate, ignoring underfill constraint, order of magnitude only). For microbumps on pitches of tens of microns, that is a real tug-of-war repeated many times a day. Using a Coffin-Manson relation with an exponent near 2, cutting the swing from 55 K to 35 K extends fatigue life by about 2.5x (author estimate).

Point 3: Gradients are worse than averages

Inside one package, the GPU core runs hottest, the HBM stacks beside it run cooler, and the substrate cooler still. Uneven temperature means uneven expansion, so the package bends into saddle, smile or frown shapes. That is the gradient problem Jun He pointed to: the average temperature looks fine while local stress delaminates low-k dielectric or chips die corners. Liquid's high heat capacity and adjustable flow narrow the gap between hotspot and surroundings, not just the peak.

Point 4: Next, coolant inside the silicon

At ECTC 2025 TSMC showed a silicon-integrated micro cooler (IMC-Si) on a 3.3x-reticle CoWoS-R package: with 40 °C water at 10 L/min it removed up to 3.4 kW (about 2.5 W/mm²), used a warpage-tolerant sealant, and passed 2,000 temperature cycles (IEEE ECTC 2025). That is a demonstration, not a confirmed Blackwell production design, but the direction is clear: as packages grow, cooling has to move into the package itself.

Open questions

• NVIDIA and TSMC have not published B200 warpage, bump pitch or reliability specs; the swing and lifetime figures here are order-of-magnitude estimates from public material constants.
• The 2024 CTE-mismatch account comes from analysts and press; NVIDIA itself only said it changed the GPU mask to improve yield.
• The OCP remarks are quoted from a third-party recap, not an official TSMC transcript.

Related videos

You Think TSMC's CoWoS is Just One Technology? (S, R, & L Explained)

The End of Overheating? Inside the 1mm "Solid-State" Fan

Where do you think AI accelerator cooling goes next: bigger cold plates, immersion, or microchannels etched into the silicon? Tell me in the comments.

Sources: NVIDIA HGX docs, SemiAnalysis, 3D InCites, IEEE ECTC 2025, OCP APAC 2026 recap, Barrack AI. Figures marked "author estimate" are my own calculations, not official numbers.
Published on The Automation Core. Videos by @TheEngineeringCore-v. Join the discussion on the PLC Facebook page.