Modern SoCs are split into chiplets that must be re-connected inside the package. The choice of interconnect — plain organic substrate, a silicon interposer, an embedded bridge — sets the achievable die-to-die pitch, bandwidth density, and cost. Select an example:

Figure 1 · Cross-section
silicon die Si interposer / bridge organic substrate PCB µbump (fine) C4 bump BGA ball LGA land

Side-by-side comparison

PackageDie-to-die mediumInterconnect pitchNeeds silicon?Rel. costDie-to-die BW densityReal-world example
MCM (organic)package-substrate tracescoarse — C4 ≈130 µm + substrate linesno$lowAMD server MCMs, RF/PA modules
2.5D interposerSi-interposer RDL + µbumpsfine — µbump ≈40–55 µm, RDL sub-µmyes (large interposer + TSVs)$$$very highGPU + HBM (TSMC CoWoS), Xilinx SSIT
EMIB (bridge)small Si bridge embedded in substratefine, but only locally at the die edgeyes (tiny bridge, no big TSVs)$$high (local)Intel Sapphire Rapids, Stratix FPGAs
LGA (board attach)— not a die-to-die method —land ≈1 mm (socket)no$$ (socket)n/aIntel LGA1700, AMD SP5 CPU sockets

Note: LGA is a board-attach style (the alternative to soldered BGA balls) — it's about how the whole package meets the motherboard, not how two dies talk. It's here because it's the other "substrate" question people ask about.