By Hamza Faisal

The chiplet era just got its rulebook for the whole package. The Universal Chiplet Interconnect Express (UCIe) Consortium has released the UCIe 4.0 specification, moving the open chiplet standard beyond simply wiring one die to another and into managing, powering and policing an entire multi-die system. The release, announced on November 12, is the biggest update to the standard since it reached 64 gigatransfers per second with version 3.0.

Chiplets — small, specialised silicon dies combined inside a single package — are now the default way to build AI accelerators, server processors and high-end mobile chips. Until now, UCIe mostly answered one question: how do two dies talk to each other? Version 4.0 answers the harder ones: how does memory reach the compute die, and how do dozens of chiplets from different suppliers stay in sync?

Memory gets its own die

The headline addition is UCIe.mem, a standardised on-package memory link between a system-on-chip and a separate memory logic die. In practice, it lets the memory controller live on its own piece of silicon instead of inside the compute die, while still working with existing memory technologies such as HBM and LPDDR. It reuses the existing UCIe physical layer and adapter, so designers do not need a second interconnect ecosystem just for memory.

The consortium says the link can reach up to 93.75 percent data efficiency on read-heavy traffic and about 2.5 terabytes per second of bandwidth per millimetre of die edge. Those are specification targets rather than results from shipping silicon, but they point at the real bottleneck in modern AI hardware: not raw compute, but getting data to the compute fast enough.

Power, trust and safety built in

UCIe 4.0 also adds an optional power-optimised profile that the consortium says delivers 30 to 80 percent greater energy efficiency at 64 gigatransfers per second, along with an enhanced bump map that raises shoreline bandwidth density by about 20 percent. Designers can trade that headroom for more bandwidth in the same package footprint, or hand the space back to other functions.

On the trust side, the specification introduces Trusted Reporting of Chiplet State, or TRoCs, which lets a system verify the reported operating condition of each individual chiplet and guard against spoofed or manipulated status information. Two-way integrity-protected management channels shield communications between chiplets from unauthorised modification, and enhanced reset management gives the system finer control over how individual dies start, stop and recover.

Automotive applications get attention too: optional error notifications and link-health monitoring feed interconnect integrity data into broader vehicle reliability and functional-safety systems.

Why it matters for AI hardware

The timing is no accident. AI accelerators are hitting the limits of single-die design, and the industry's answer is to stitch together compute, memory and input-output dies from multiple suppliers in one package. That only works if the dies agree on how to behave as a system — which is exactly what UCIe 4.0 standardises. The consortium counts more than 150 members, with board companies including AMD, Arm, ASE, Alibaba, Google Cloud, Intel, Meta, Microsoft, Qualcomm, Samsung and TSMC.

The specification stays fully backward compatible, and the top signalling rate is unchanged at 64 gigatransfers per second — this release is about system behaviour, not raw speed. The specification and an accompanying white paper are available by request; no vendor has yet announced products built on the new features. But the direction is clear: the chiplet is no longer just a packaging trick. It is becoming a system, with a standard to match.