Summary

Close the subset of performance-phase floor gaps that don't require testbed access:

1. recipes/overlays/gb200-eks-ubuntu-inference-dynamo.yaml — add an inference-perf block using the same placeholder smoke thresholds as h100-eks-ubuntu-inference-dynamo.yaml:65-79. GB200 outperforms H100, so the H100 floors (inference-throughput >= 5000, inference-ttft-p99 <= 200) are trivially satisfied — they assert "the system runs," not "GB200 is fast enough." Comment that the values are placeholders pending GB200-tuned numbers.
2. recipes/overlays/h100-eks-ubuntu-inference-nim.yaml — add the same placeholder inference-perf block (NIM uses the same check name and threshold set).
3. pkg/recipe/validation_phase_floor_test.go::requiresPerformance — return false when accelerator is unset. This exempts accelerator-unbound intermediates (aks-training, eks-training, gke-cos-training, oke-training) from the strict-mode perf floor because the threshold is accelerator-specific and only meaningful on concrete leaves.

Motivation / Context

AICR_VALIDATION_FLOOR_STRICT=1 go test ./pkg/recipe/... -run TestOverlayValidationPhaseFloor flags 13 overlays today. PR #1001 closes the deployment-phase gap (separate axis). This issue closes 6 of the 13 remaining perf warnings without needing testbeds:

The remaining 7 are testbed-blocked AKS / OCI / LKE leaves, tracked in dedicated issues.

Design notes

• The two Dynamo / NIM placeholder blocks follow the precedent set by h100-eks-ubuntu-inference-dynamo.yaml — loose smoke-test floors with a comment explaining they're not tight perf gates. Tightening to empirical thresholds is a separate per-accelerator workload-tuning effort.
• The requiresPerformance refinement is a 3-line rule change: if c.Accelerator == "" || c.Accelerator == "any" { return false }. The rationale: an intent-level intermediate without accelerator binding has no meaningful perf threshold; concrete leaves (h100-eks-training ≥ 300 GB/s, gb200-eks-training ≥ 40 NET / 500 NVLS GB/s) carry their own threshold via per-phase replace. The floor test was over-broad for these intermediates.

Done when

• go test ./pkg/recipe/... -run TestOverlayValidationPhaseFloor -v continues to pass in default mode.
• AICR_VALIDATION_FLOOR_STRICT=1 go test ./pkg/recipe/... -run TestOverlayValidationPhaseFloor -v failures drop from 13 to 7 (testbed-blocked only).
• The 6 newly-closed cases never reappear in either mode.

Related

• Close deployment-phase validation coverage gaps for accelerator-bound GPU recipes #969 — deployment-validation coverage (rescoped, separate axis)
• feat(recipe): deliver deployment-phase floor at per-accelerator wildcards #1001 — PR closing Close deployment-phase validation coverage gaps for accelerator-bound GPU recipes #969 (deployment phase)
• AKS / OKE / LKE testbed-blocked perf issues — filed separately
• feat(recipe): per-field union merge for validation phase checks #1000 — per-field union merge for validation checks (orthogonal merge-semantics fix)