Home » Uncategorized » UCIe 4.0 Spec Adds a Standard Memory Link for Chiplets

UCIe 4.0 Spec Adds a Standard Memory Link for Chiplets

A new rulebook lets processors hand memory control to a separate small die, freeing space at the chip edge. It is a published standard, not a product, and no silicon…

Close-up of a patterned silicon wafer, illustrating the UCIe 4.0 chiplet standard

Where should a processor keep its memory controller? Until now, chip designers have answered that question their own way. On November 12, the UCIe Consortium published a UCIe 4.0 chiplet specification that proposes one shared answer for memory links inside a package.

Here is what is verified. The consortium’s own release confirms the date, the new memory protocol and the headline figures. Two independent outlets, Jon Peddie Research and XenoSpectrum, report the same numbers. This is a published standard. It is not a chip you can buy, and no source gives a date for products.

What the UCIe 4.0 Chiplet Spec Changes

A chiplet is a small piece of silicon that does one job. Several chiplets sit together in one package and act like a single processor. UCIe, short for Universal Chiplet Interconnect Express, is the set of rules for how those pieces talk to each other.

The big addition in 4.0 is UCIe.mem. It defines a standard link between a compute chip and a separate memory logic chiplet, using existing memory technologies. Jon Peddie Research reports that this moves memory control off the compute die. It says the logic chiplet can manage HBM (high-bandwidth memory, the stacked memory used by AI accelerators), LPDDR (low-power memory common in phones) or future memory types.

XenoSpectrum adds that the link carries requests and responses, not raw memory signals. It reuses the existing UCIe physical layer, adapter and compliance testing. That matters for designers, because they can reuse work they have already done.

The consortium describes the shift in plain terms. UCIe is moving from defining how single chiplets connect to defining how several chiplets work together as a system, said consortium president Cheolmin Park of Samsung.

The Numbers Behind the Release

The consortium lists these figures for the new memory link and related changes:

  • Data efficiency as high as 93.75%, meaning the share of link capacity that carries useful data.
  • Die-edge bandwidth density of about 2.5 TB/s per millimeter. The die edge, sometimes called shoreline, is the limited border of a chip where connections can leave.
  • An optional power profile for 64 GT/s links that the consortium says gives up to 30 to 80 percent better energy efficiency. GT/s means gigatransfers per second, a measure of link speed.
  • A revised bump map for the short-reach UCIe-S variant at 48 and 64 GT/s, with about 20 percent more bandwidth along the edge and 10 percent more per unit of area.

Jon Peddie Research did its own arithmetic. At 2.5 TB/s per millimeter, 10 millimeters of edge would give 25 TB/s in theory. It stresses that real bandwidth depends on the package, the controller and the memory used. XenoSpectrum makes the same point: the density figure is not the effective memory bandwidth of a whole chip.

Management and Security Updates

The spec also adds features for packages that mix chiplets from several vendors. Trusted Reporting of Chiplet State lets a root of trust in the package check each chiplet’s identity and state. Management channels now have integrity protection, and reset handling is more consistent. Optional error notifications and link-health monitoring target automotive reliability and functional safety.

The consortium says 4.0 is fully backward compatible with earlier UCIe versions. It is available to the public by request. The group, which counts more than 150 member companies, is led by firms including ASE, Alibaba, AMD, Arm, Google Cloud, Intel, Meta, NVIDIA, Qualcomm, Samsung and TSMC.

What Is Still Unproven

Start with the power claim. It is a link-level figure, not a system-level one. Jon Peddie Research notes the savings need conditions to hold: both dies on similar process generations, shared voltage rails and a distance of about 5 millimeters or less. The consortium’s release does not say what baseline the 30 to 80 percent is measured against.

Nothing here is measured silicon. Neither outlet found independent measurements, and the consortium sells no chips. Jon Peddie Research says no commercial deployment of UCIe 4.0 memory links is established. Its analyst view is that the spec offers a path, not an inflection point, and it wants test silicon, interoperability results and production commitments first. It points to 2027 and 2028 as the years to watch.

For scale, the same report cites AMD’s Versal RF-series, which uses native UCIe 1.1 with production chiplets expected in late 2027. That is an older version of the standard, and it shows how slowly spec releases turn into shipping parts. Verification across chiplets from different makers is also a known hard problem, according to the report’s named experts.

Why It Matters to You

You will not see UCIe 4.0 on a spec sheet this year. If designers adopt it, the likely payoff is processors that pack more memory bandwidth into the same package, or that use less power moving data. That could eventually help AI servers and phones with big on-chip workloads.

Expect any effect on price, battery life or speed to arrive after 2027, and only if vendors build to the standard. AMD’s own move toward 256-core server chips shows why the memory edge of a chip is getting crowded.

One thing to watch next: the first compliance results from two different vendors’ chiplets talking over UCIe.mem.

Sources

  • UCIe Consortium, “UCIe Consortium Advances from Chiplet Connectivity to System-Level Integration with UCIe 4.0 Specification,” November 12, 2026.
  • I-Connect007, “UCIe Consortium Releases 4.0 Specification for System-Level Integration” (syndicated consortium release), November 13, 2026.
  • Jon Peddie Research, “UCIe 4.0 brings memory into chiplets,” November 16, 2026.
  • XenoSpectrum, “UCIe 4.0 Released: New Memory Interface Claims 94% Transfer Efficiency, Up to 80% Lower Link Power,” November 13, 2026.

Related Reading