TheDZ is currently undergoing a market data infrastructure update. Some live features may be unavailable.

SPECULATIVE REFERENCE DESIGN — NOT AN INTEL OR CORINTIS PRODUCT

Pablo’s Chip

XBM-14A / CX-1

A 3.2 kW accelerator package built on Intel 14A compute tiles, EMIB-T bridges, and 12 stacks of Intel’s patented cross-batch memory — cooled by a three-tier Corintis microfluidic system that treats the memory base die, not the logic, as the hardest thermal problem.

Package
120 × 180 mm, glass substrate
Compute
4 × 14A tile, High-NA EUV
Memory
12 × XBM 16-high, 576 GB
Cooling
Corintis 3-tier co-packaged
Earliest plausible
2030+

Cross-section

Coolant 45 °C inDRAM tier ≤ 85 °CLogic junction ≤ 105 °CBase-die SerDes hotspot
In — 45 °C facility water, no chillerOut — 60 °C, heat-recoverable
  • Tier 1 · zoned microfluidic lid16 independently solved flow zones · 80 µm pitch Cu, Glacierware-generated

    The lid is not a cold plate with one flow path. Sixteen zones are solved independently so the coolant arrives cold where the flux is highest, instead of arriving pre-heated after crossing the compute tiles. Loop A reaches the DRAM stacks cold-first — the 85 °C refresh ceiling is tighter than the logic ceiling, so it gets first claim on the coldest water.

  • Tier 2 · backside channels in 14A siliconCoolant ~50 µm from junction

    Channels etched into the back of the compute die put coolant roughly 50 µm from the junction. This is what makes the 400 W/cm² matrix-engine hotspots survivable: at that distance the silicon between heat source and fluid stops being the dominant thermal resistance.

  • 14A tile 0600 W · 400 W/cm² pk

    One of four Intel 14A compute tiles on High-NA EUV. 600 W each, 2,400 W across the four. Average flux is a manageable 120 W/cm²; the problem is the peak — matrix-engine hotspots hit 400 W/cm², which is where Tier 2 zone-solving stops being a nicety.

  • 14A tile 1600 W · 400 W/cm² pk

    Identical to tile 0. Tiles are bridged with EMIB-T at 25 µm bump pitch rather than sitting on a full-surface silicon interposer — cheaper and shorter, but it removes a lateral heat-spreading path the package would otherwise have had for free.

  • × 16 tiers XBM · BEOL 1T1C DRAM≤ 85 °C · 12 stacks, 576 GB total

    Cross-batch memory built as 1T1C DRAM in the back end of line — thin-film transistors in the metal stack rather than front-end silicon. Twelve stacks, sixteen high, 420 W in aggregate at a mild 18 W/cm². The tier is thermally easy and structurally hard: retention, leakage and repair behaviour all differ from conventional DRAM at this height.

  • UCIe base die32 GT/s SerDes · ~90 W in 40 mm² · 1,080 W across 12

    The whole reason this package is interesting. Serialising the memory interface concentrates I/O power into the base die instead of spreading it across PHY on the SoC edge: ~90 W in 40 mm² is 225 W/cm², buried under sixteen DRAM tiers. Top-side cooling cannot reach it, which is what forces Tier 3 to exist.

  • Tier 3 · riserLateral extraction from the base die

    The load-bearing invention, and the only part of this design with no public demonstration behind it. Pulling heat sideways out of a base die buried under sixteen tiers is a research problem, not a product. Everything above it in this stack is incremental engineering; if the riser does not work, the package does not exist.

  • Glass substrate · EMIB-T25 µm bump pitch · 120 × 180 mm, 12+ reticle-class silicon area

    A 120 × 180 mm glass substrate carrying more than twelve reticles of silicon. Glass buys dimensional stability at this size; dropping the interposer for bridges buys z-height, which is what leaves a stack alley wide enough for the riser to fit at all.

  • OAM-class carrierQuick-disconnect coolant fittings at card edge

    Coolant enters at 45 °C facility water — no chiller — and leaves at 60 °C, warm enough to be worth recovering. About 6 L/min for the whole ~4.2 kW package. The fittings are also the commercial problem: liquid inside a $40k package is a warranty question long before it is a physics question.

Each layer carries a design decision, not just a label. Click or tab to any block in the drawing to open it. Layer thicknesses are schematic, not to scale. Colour encodes steady-state temperature class, per the legend above.

Thermal budget

Modeled — illustrative, not simulated

ZonePowerFluxCeilingCooling tierMargin
4 × 14A compute tile2,400 W120 W/cm² avg105 °C TjTier 1 + 2wide
Matrix-engine hotspots400 W/cm² pk105 °C TjTier 2, zone-solvedtight
12 × XBM DRAM tiers420 W18 W/cm²85 °CTier 1, Loop A cold-firstwide
12 × UCIe base die1,080 W225 W/cm²95 °CTier 3 riser onlycritical
Substrate, VR, leakage~300 WlowTier 1 spilloverwide
Package total~4.2 kW45 °C in / 60 °C out~6 L/min

The inversion

Why XBM changes the cooling problem

HBM4 packageXBM packageConsequence
Interface~2,048-bit parallel, slowSerial UCIe, 32 GT/sInterface power concentrates instead of spreading
Where I/O power sitsSplit across PHY on the SoC edgeIn the memory base dieNew hotspot appears under the memory
InterposerSilicon, full-surfaceNone — bridges onlyLoses a large lateral heat-spreading path
Z-heightTaller, stiffenersShorter, no stiffenersRoom for a riser in the stack alley
Thermal binding constraintDRAM refresh at 85 °CBase-die SerDes at 95 °CTop-side cooling stops being sufficient

What kills it

Ranked by how fast it is fatal

Fatal if unsolved

The riser is fiction

Tier 3 is the only part of this design with no public demonstration behind it. Lateral extraction from a base die buried under sixteen tiers is a research problem, not a product. Everything above it is incremental; this is the load-bearing invention.

Fatal if unsolved

Coolant inside a $40k package

Warranty, not physics. A hyperscaler designing its own silicon absorbs leak risk on its own fleet. A merchant foundry selling this to third parties absorbs it on someone else’s, and gets an RMA queue for the privilege.

Schedule risk

Two timelines that barely touch

14A hits volume in 2028. XBM is a December 2024 patent filing with commercialization talked about at 2030 or later. A package needing both is a 2030-at-the-earliest object, and only if Intel actually re-enters DRAM.

Schedule risk

BEOL DRAM is unproven at 16-high

Thin-film transistors in the metal stack change retention, leakage, and repair behaviour versus front-end silicon. Intel’s own answer is built-in self-repair and spare channels — which is a yield admission, and yield admissions carry thermal margin costs.

Structural

The board seat cuts both ways

Lip-Bu Tan sits on Corintis’ board personally. That aligns the thinking and complicates the paperwork: any Intel–Corintis commercial agreement runs through recusal and arm’s-length review before it runs through engineering.

Structural

Intel may just build it

Intel has its own packaging thermal program and has already shown segmented heat spreaders for kilowatt-class parts. The buy case rests on Corintis’ flow-optimization software and copper microfabrication, not on the physics, which Intel understands perfectly well.

Speculative engineering exercise. Intel has announced no XBM product, no Corintis relationship, and no co-packaged cooling offering; XBM exists as a patent application published 2 July 2026. Public anchors used: Intel EMIB-T scaling and 120 × 180 mm package roadmap; Intel 14A PDK 0.9 targeted October 2026 with volume in 2028; the XBM patent’s 1T1C BEOL DRAM, 0.5–5 GB dies, 8–16-high stacks, and 32 GT/s UCIe base-die I/O; and the Microsoft–Corintis in-chip microfluidic result. Every power, flux, flow, and temperature figure above is the author’s estimate.