An Unnamed x86 CPU Has Run Legacy-Free for a Year

A request to document 64-bit-only behaviour for the Linux kernel describes an x86 part that dropped the modes Intel decided to keep, and says it has been in service for over a year.

Intel published the first x86S specification three years ago and stopped work on it in mid-2024, leaving the cleanup of x86's boot-time and 32-bit baggage as an abandoned proposal. A message relayed to the Linux kernel mailing list says somebody else carried on with it and has been running the result in production for more than a year.

A 64-bit-only x86 that nobody will name

The implementation described in that message runs 64-bit protected mode only, requires five-level paging, and carries both FRED and APX, with Legacy Mode and Compatibility Mode removed outright. The request was to have the resulting behaviour documented for the kernel, and it arrived through Christian Ludloff, the x86 architecture specialist behind sandpile.org, relaying it on behalf of a company he does not name, as Phoronix set out from the mailing list post. The poster's own line is that the design "has been used at scale for >1 year".

Dropping Compatibility Mode is the part that goes further than Intel's draft ever did. Compatibility Mode is what lets a 64-bit operating system run 32-bit user-space binaries, so a chip without it runs no 32-bit code at all, and losing Legacy Mode means the reset and boot path has to be rebuilt as well, since a stock x86 still comes out of reset in real mode. That combination is appliance or datacenter silicon by construction. It cannot boot a retail Windows install, and it would not run a 32-bit game even under emulation without a software layer doing the work.

The ordering is what makes it interesting rather than the mystery. Intel published the FRED specification five years ago and has only now shipped its first parts carrying it, with AMD expected to follow; the APX specification is three years old and neither vendor has shipped it yet. An unnamed company has both in silicon that it says has been in service since before either. The same outfit earlier disclosed 16-tile and 32-tile AMX implementations, wider than the Advanced Matrix Extensions in Intel's current Xeons.

Who it is remains open. Phoronix raises Zhaoxin and Hygon as the obvious candidates and floats RosaicLabs, a start-up reported to have obtained Atom designs from Intel, while noting that bolting 32-tile AMX, APX, FRED and a legacy-free mode structure onto an Atom-derived core would be a considerable leap. None of it is verifiable from outside: there is no product, no die shot, and no part number, only a documentation request and the claims attached to it.

Nova Lake starts with the smaller die

Intel's first Nova Lake-S desktop chips enter mass production at the end of this year and go on sale in the first quarter of 2027, with the 28-core single-compute-tile models arriving first and the dual-tile parts reaching up to 52 cores later in 2027, on a partner presentation slide posted to X and translated by HotHardware. The same slide has Razor Lake possibly following in late 2027 and keeps LGA1954 on LPDDR5 across more than one CPU generation. It is a leaked vendor deck rather than an Intel statement, so the dates are reported rather than committed.

Leading with the smaller part is backwards for Intel, which normally launches its halo SKU first and backfills underneath it. Two compute tiles on a consumer desktop CPU is something Intel has not shipped before, and the single-tile parts have to carry the gaming argument against AMD's 3D V-Cache chips in the meantime, which is what Nova Lake's large last-level cache is there for. HotHardware's own caution on the 52-core parts is worth keeping: core count does not produce multi-core scaling on its own, and a desktop chip that wide can end up waiting on its memory interface rather than on cores.

AMD looks set to fill a hole in AM5's entry tier

A standard Ryzen 5 7500 is reportedly in preparation for AM5: six Zen 4 cores and twelve threads at a 3.7GHz base and up to 5.0GHz boost, 38MB of total cache from 32MB of L3 and 6MB of L2, a 65W TDP, integrated graphics, and a boxed cooler, from specifications shared by Roland Quandt and written up by KitGuru. The boost clock matches the existing tray-only 7500F, so the display output and the cooler in the box are what separate them, along with a reported price near 230 euros. AMD has not confirmed the chip or a release window.

That gap is real if the part lands. The 7500F has no integrated graphics and ships without a cooler, which makes it awkward as the cheap way onto AM5 for anyone assembling a machine before the graphics card arrives, or troubleshooting one without a spare. The value of the SKU rests entirely on where retail pricing settles against the 7500X3D above it.

None of this changes an order placed this week. APX is the line worth watching: it is the first x86 extension in years that needs a recompile before it does anything, so the interesting question once AMD and Intel both ship it is which distributions and which game runtimes actually build for it.

Sources