UFI (Ultra Express Fabric)

NULL - deliberately, and this record exists to flag an unfilled hole rather than to assert anything. No primary UEFI Forum specification release for UFI is retrievable as of the access date. This was re-checked on 2026-10-03 with a wider search than the first pass: uefi.org/specifications lists ACPI 6.6, UEFI 2.11, PI 1.9 and the PI distribution packaging spec, and NO UFI release; uefi.org/workinggroups lists exactly five working groups (ASWG, ICWG, PIWG, USWG, UTWG), none of them a UFI or fabric group; and repeated exact-phrase web searches for 'Ultra Express Fabric' return no UEFI Forum document at all. So the absence is now positively established against the publisher's own site, not merely unfound. Per this repo's contract a wrong confident record is worse than an open question, so no version, no bandwidth, and no link count are asserted. Do not populate this record from secondary summaries.

OPEN ITEM, deliberately left as nulls - but the provenance is now stronger and the open question is now precisely stated. VERIFIED ABSENT AS OF 2026-10-03: (1) uefi.org/specifications carries no UFI entry; its latest releases are UEFI 2.11 (December 2024), PI 1.9, ACPI 6.6 (May 2025); (2) uefi.org/workinggroups enumerates all five UEFI Forum working groups and none concerns UFI or any fabric; (3) exact-phrase search for 'Ultra Express Fabric' returns zero UEFI Forum documents. The first pass of this record cited only the specifications index; the working-groups check was added because a spec that exists but is not yet released would plausibly still have a chartered working group, and it does not. That is a stronger negative result than the original. WHAT IS KNOWN FROM CONTEXT ONLY, and therefore NOT recorded in any typed field: that UFI is a UEFI Forum initiative intended as a memory-and-I/O-expansion fabric for platforms, built on PCIe and CXL. If you have a primary UEFI Forum document that contradicts this record, it should supersede it. INFERENCE RELEVANCE IF IT LANDS: on the same reasoning already applied to cxl-2-0 through cxl-4-0 in this repo, a memory-expansion fabric tier is a CAPACITY lever for large-batch decode and KV-cache headroom, not a tensor-parallel lever - it would not change collective bandwidth inside a node. It would matter for exactly the workloads this repo tracks as memory-bound: growing KV cache beyond HBM capacity without paying for a faster local memory tier. THAT SAID, the practical near-term answer to that problem is already covered by cxl-4-0, so UFI's marginal value OVER CXL is the open question, not its relevance. TO RESOLVE: locate a UEFI Forum specification release or a UFI specification PDF, then fill version, bandwidth_gbps (per lane, per direction, with the direction convention stated in bandwidth_basis), topology, and scale_up / scale_out. Until then this record should not be cited by any benchmark or gotcha. KEPT AT status draft / confidence 0.2 deliberately: a low-confidence record that is honest about being empty is more useful than a filled-in guess, and the null fields here are a to-do marker, not an oversight. 2026-10-10: set null_bandwidth=verified-negative. The record's own bandwidth_basis already says this; the field only names the case. No number changed. The basis says the absence was re-checked against uefi.org/specifications and uefi.org/workinggroups on 2026-10-03 and no UFI release exists. A fetched publisher site that contains no spec is verified-negative, not 'not yet looked'.