Runtime map
A source-linked inventory of what Alpha currently maintains, what the compiler emits, and which physical devices have actually closed an evidence path.
Evidence boundary
This page is bound to e2ff6874. It includes committed source and retained evidence only; working-tree experiments are excluded until they become part of a deployable revision.
A retained device receipt closes the stated path.
The compiler can produce this host artifact.
The owner exists and is maintained; broader execution is not implied.
A checked boundary records why the current artifact is incompatible.
Program roots
Systems are actual programs under systems/, not compiler internals or hardware packages.
The historical Alpha-ER model family with its model-specific training, inference, and SM86 composition roots.
The named 60M-class transformer system, including its model, learner, arena, launch policy, artifact identity, and physical attestation.
Execution layers
A host output, an accelerator architecture, and a physical device are different claims. The table keeps those layers separate.
| Target | Layer | Current state | Identity and boundary | Evidence |
|---|---|---|---|---|
| x86_64-linuxELF64 / x86-64 / Linux | Host output | Compiler emits | The default native host target and the host side of the current Coppelius training artifact. A host target is not an accelerator qualification. GPU execution still depends on the Linux/NVIDIA platform and a compatible device image. | |
| aarch64-linuxELF64 / AArch64 / Linux | Host output | Emits and runs | The compiler emits this target through the direct lane; emitted programs were physically executed on the DGX Spark host. Direct lane only. The evaluator-carrying artifact, native object output, and tune lane remain x86-64-only. | |
| sm86-systemNVIDIA SM86 / compute capability 8.6 | Accelerator architecture | Named, not CLI-emittable | Owns the architecture, command formats, kernels, and cooperative realization used by the current SM86 device profiles. The compiler knows this target name but refuses it through the generic --target selector: the complete system path is assembled by system-owned build and qualification roots. | |
| GeForce RTX 3070GA104 / SM86 / 8,192 MiB VRAM | Physical device | Physically qualified | Coppelius default physical binding and the constrained SM86 qualification device. Qualification is evidence for the frozen sm86-compat profile and its stated resource bounds—not a claim that every Alpha program runs on this device. | |
| GeForce RTX 3090GA102 / SM86 / 24,576 MiB VRAM | Physical device | Physically qualified | The unchanged Coppelius sm86-compat ELF passed fresh and resumed training on this device. This is a second physical witness for the same compatibility artifact, not a separate 3090 target or a larger-memory profile. | |
| DGX Spark / NVIDIA GB1020-core AArch64 / Blackwell 12.1 / 130,596,048,896 B unified memory | Physical device | Host yes · accelerator no | A measured and qualified AArch64 build/execution host for Alpha's direct native target. Its GB10 accelerator is not admitted by the SM86 capability lens. Current SM86 artifacts do not run there, and Alpha records no GB10 QMD, pushbuffer, or GPFIFO support. |
Relationships
Follow a row from program ownership to host output, accelerator contract, and the physical evidence that exists today.
| System | Host | Accelerator | Physical devices |
|---|---|---|---|
| Alpha-ERMaintained program root | x86_64-linuxHost/runtime families | SM86Admitted-plan composition present | No system-specific physical qualification claimed here |
| CoppeliusPhysically qualified system | x86_64-linuxStatic host ELF | sm86-compatOne immutable compatibility artifact | RTX 3070RTX 3090Separate physical receipts |
| Compiler platformNot a model system | aarch64-linuxDirect lane; physically executed | GB10 / Blackwell 12.1Outside the SM86 architecture | DGX SparkHost qualified; accelerator not admitted |