v10 unification: OnDieFabric inheritance base + endpoint-count naming reconciliation

[Ravenwater]

Goal

Extend the v9 unification (#215) by collapsing the 6 *OnDieFabric classes (CPU/GPU/NPU/CGRA/DPU/TPU) into an inheritance hierarchy: a shared OnDieFabric base in compute_block_common carries the 7 shared fields + optional mesh dims + optional confidence; per-block-kind subclasses contribute only the architecture-specific topology enum.

Reconciles the endpoint-count naming inconsistency at the same time: CPU's stop_count and GPU's controller_count are renamed to unit_count (matches the existing convention used by NPU/CGRA/DPU/TPU; per-arch field description documents what a ""unit"" means in each context).

Why now

After the v9 sprint (#215) closed, compute_block_common hosts 4 re-exported primitives + 2 unified types (TheoreticalPerformance + ThermalProfile). The v9 paper exercise flagged OnDieFabric as the v10 candidate:

OnDieFabric unification: 3 of 6 classes (NPU/CGRA/DPU) are byte-identical (10 fields). TPU drops mesh dims; CPU uses stop_count; GPU uses controller_count for the same conceptual ""endpoint count"" field. Naming reconciliation + extension model needed. Defer to v10.

Why inheritance (vs alias)

Unlike v8/v9 where the 5 (resp. 5) target classes were byte-identical and collapsed via aliases, OnDieFabric is NOT byte-identical across the 6 block kinds because each carries its own topology: <Arch>NoCTopology enum. The topology enums are genuinely architecture-specific (e.g., CPU's infinity_fabric is meaningless to a TPU; DPU's aie_mesh is meaningless to a CPU). Preserving them is correct.

The inheritance approach gives the best of both worlds:

• Shared fields defined once (in base)
• Per-block-kind enum validation preserved
• isinstance(x, NPUOnDieFabric) AND isinstance(x, OnDieFabric) both work
• Each per-kind subclass shrinks from ~30 LOC to ~5 LOC

Survey: 6 OnDieFabric classes today

Class	Shared 7 fields	mesh_cols/rows	confidence	endpoint name
NPUOnDieFabric	YES	YES	YES	unit_count
CGRAOnDieFabric	YES	YES	YES	unit_count
DPUOnDieFabric	YES	YES	YES	unit_count
TPUOnDieFabric	YES	no	YES	unit_count
CPUOnDieFabric	YES	no	no	stop_count
GPUOnDieFabric	YES	no	no	controller_count

Shared 7 fields (all 6 have these):
bisection_bandwidth_gbps, flit_size_bytes, hop_latency_ns, pj_per_flit_per_hop, routing_distance_factor, topology (per-kind enum), endpoint-count-field.

Topology enums (stay separate)

Each block kind has its own topology enum with genuinely different values:

• CPU: ring / double_ring / io_die_plus_ccd / infinity_fabric
• GPU: crossbar / ring / hierarchical
• NPU: dataflow_ring / systolic / crossbar / shared / partitioned
• CGRA: crossbar
• DPU: aie_mesh / crossbar
• TPU: crossbar / multi_crossbar

Only ""crossbar"" appears in multiple enums. Unifying into a single enum would lose architectural specificity. Topology enums stay per-block-kind.

YAML impact

3 YAMLs reference the to-be-renamed fields:

• intel/intel_core_i7_12700k.yaml (CPU stop_count -> unit_count)
• nvidia/jetson_agx_orin_64gb.yaml (GPU controller_count -> unit_count)
• nvidia/jetson_agx_thor_128gb.yaml (GPU controller_count -> unit_count)

Same migration as any field rename: PR 3 updates these 3 YAMLs atomically with the schema change.

Sprint shape (3 PRs, mirror of v8 unification + v9 unification pattern)

1. PR 1 (graphs docs) -- paper exercise at docs/designs/v10-on-die-fabric-unification.md. Audits all 6 *OnDieFabric classes side-by-side, justifies the inheritance approach (vs alias), documents the endpoint-count rename + YAML migration plan.

2. PR 2 (embodied-schemas) -- add OnDieFabric base class to compute_block_common.py with 7 shared fields + optional mesh_cols / mesh_rows / confidence. Additive only: existing per-kind classes unchanged.

3. PR 3 (embodied-schemas) -- migrate all 6 per-kind classes to inherit from OnDieFabric base. CPU and GPU also get the stop_count -> unit_count and controller_count -> unit_count rename + 3 YAML updates in the same commit.

Expected diff: net -150 to -200 LOC of duplicated field definitions removed; 3 YAML field renames.

Out of scope (defer to v11)

• has_external_dram vs has_host_dram naming reconciliation (touches more SKU YAMLs; semantic difference matters -- chip-attached HBM vs host-bus PCIe DRAM)
• KPUNoCSpec (oldest module; pre-pattern; doubly-purposed)
• Topology enum unification (intentionally architecture-specific)

Refs

• v8 unification: graphs#208 (closed)
• v9 unification: graphs#215 (closed)
• v9 paper exercise flagged this as v10 candidate:
  graphs/docs/designs/v9-thermal-profile-unification.md (risk TASK-2026-014: Phase 0.5 — PowerMeter abstraction + composition-test CI gate #4 / summary table)