diff --git a/docs/user/validation.md b/docs/user/validation.md
index 4619b7df0..02d203a02 100644
--- a/docs/user/validation.md
+++ b/docs/user/validation.md
@@ -40,9 +40,21 @@ any phase. If pre-flight fails, no validator Jobs are deployed.
 
 ## Training performance validation
 
-Training performance runs the `nccl-all-reduce-bw` check — a Kubeflow `TrainJob`
-that runs the canonical `all_reduce_perf` benchmark across all GPU nodes and
-measures aggregate bus bandwidth.
+Training performance runs an NCCL all-reduce benchmark — a Kubeflow `TrainJob`
+that runs `all_reduce_perf` across GPU nodes and measures aggregate bus
+bandwidth. Three check variants are available; the recipe picks the one (or
+ones) that match the target fabric:
+
+| Check | Transport | When it's selected |
+|---|---|---|
+| `nccl-all-reduce-bw` | Auto-detect (whatever NCCL picks) | Default for H100 on EKS/GKE, and for GB200/B200 on non-EKS services. Preserves the pre-variant behavior. |
+| `nccl-all-reduce-bw-net` | NET (EFA on EKS) | GB200 + EKS. Asserts EFA actually carried traffic — catches silent fallback to Socket when the NVIDIA driver is missing `NVreg_GrdmaPciTopoCheckOverride=1`. |
+| `nccl-all-reduce-bw-nvls` | NVLS (MNNVL across an NVL72 IMEX domain) | GB200 + EKS. Asserts the NVLS communicator actually initialized — catches silent fallback to EFA when the IMEX domain is misconfigured. |
+
+GB200/EKS recipes (both `training` and `inference` intents) enable `-net` and
+`-nvls` together rather than the auto-detect variant, because those nodes
+expose two inter-node fabrics simultaneously and a single auto-detect test
+would only exercise one of them.
 
 ```bash
 # Capture snapshot, generate training recipe, validate the performance phase.
@@ -55,24 +67,26 @@ aicr recipe --service eks --accelerator h100 --os ubuntu \
 aicr validate --recipe recipe.yaml --snapshot snapshot.yaml --phase performance
 ```
 
-The generated recipe lists `nccl-all-reduce-bw` under
+The generated recipe lists the selected variant(s) under
 `validation.performance.checks` with a platform-tuned bandwidth constraint
-(example: `>= 300 GB/s` for H100 + EFA).
+(example: `>= 300 GB/s` for H100 + EFA; `>= 40 GB/s` NET and `>= 500 GB/s`
+NVLS for GB200 + EFA, each sized for a 2-node pair).
 
-Expected flow (~5–10 min): readiness pre-flight → deploy `TrainingRuntime` +
-`TrainJob` in `aicr-validation` → worker pods reach `Running` → run
-`all_reduce_perf` → parse peak bus bandwidth → compare to recipe constraint
-(10 % tolerance) → cleanup.
+Expected flow (~5–10 min per variant): readiness pre-flight → deploy
+`TrainingRuntime` + `TrainJob` in `aicr-validation` → worker pods reach
+`Running` → run `all_reduce_perf` → parse peak bus bandwidth → verify the
+intended transport actually carried traffic (for `-net` / `-nvls`) → compare
+to recipe constraint (10 % tolerance) → cleanup.
 
 A passing CTRF entry:
 
 ```json
 {
-  "name": "nccl-all-reduce-bw",
+  "name": "nccl-all-reduce-bw-net",
   "status": "passed",
   "suite": ["performance"],
   "stdout": [
-    "NCCL All Reduce bandwidth: <actual> GB/s",
+    "NCCL All Reduce bandwidth (nccl-all-reduce-bw-net): <actual> GB/s",
     "Constraint: >= <threshold> → true"
   ]
 }
diff --git a/recipes/overlays/gb200-eks-inference.yaml b/recipes/overlays/gb200-eks-inference.yaml
index 205fb951e..a6d9dac0d 100644
--- a/recipes/overlays/gb200-eks-inference.yaml
+++ b/recipes/overlays/gb200-eks-inference.yaml
@@ -33,13 +33,30 @@ spec:
       value: ">= 1.32.4"
 
   componentRefs:
+    # GB200+EKS GPU Operator prerequisites. Keep in sync with the same block
+    # in gb200-eks-training.yaml. The NVreg flag and cdi+gdrcopy are needed
+    # for multi-node inference that crosses EFA — tensor-parallel serving
+    # over the network has the same dma-buf attach requirement as training
+    # all-reduce. The mixin system can't extend componentRefs already
+    # declared in the inheritance chain, so this lives per-leaf for now.
     - name: gpu-operator
       type: Helm
+      manifestFiles:
+        - components/gpu-operator/manifests/kernel-module-params.yaml
       dependencyRefs:
         - nfd
         - cert-manager
         - kube-prometheus-stack
         - nodewright-customizations
+      overrides:
+        cdi:
+          enabled: true
+        gdrcopy:
+          enabled: true
+        driver:
+          version: 580.126.20
+          kernelModuleConfig:
+            name: nvidia-kernel-module-params
 
     # GB200 uses nodewright no-op: the H100 tuning packages (nvidia-setup,
     # nvidia-tuned) are not compatible with GB200's ARM64 host CPU and
@@ -55,3 +72,20 @@ spec:
         intent: inference
       dependencyRefs:
         - nodewright-operator
+
+  # NCCL fabric health checks. NET exercises EFA (inter-node), NVLS exercises
+  # MNNVL (intra-NVL72). Both matter for multi-node inference that spans the
+  # fabric (tensor-parallel serving, MoE expert parallelism); single-node
+  # deployments hit the WorkerCount < 2 skip path gracefully. Thresholds sized
+  # for a 2-node GB200 pair — will be raised once production NVL72 data is
+  # available.
+  validation:
+    performance:
+      checks:
+        - nccl-all-reduce-bw-net
+        - nccl-all-reduce-bw-nvls
+      constraints:
+        - name: nccl-all-reduce-bw-net
+          value: ">= 40"
+        - name: nccl-all-reduce-bw-nvls
+          value: ">= 500"
diff --git a/recipes/overlays/gb200-eks-training.yaml b/recipes/overlays/gb200-eks-training.yaml
index bce387c14..00f201b96 100644
--- a/recipes/overlays/gb200-eks-training.yaml
+++ b/recipes/overlays/gb200-eks-training.yaml
@@ -33,9 +33,19 @@ spec:
       value: ">= 1.32.4"
 
   componentRefs:
-    # GB200-specific GPU Operator overrides (inherits valuesFile from eks-training)
+    # GB200+EKS GPU Operator prerequisites. Keep in sync with the same block
+    # in gb200-eks-inference.yaml — the mixin system can't extend componentRefs
+    # already declared in the inheritance chain (gpu-operator comes from
+    # eks-training / eks-inference), so this lives per-leaf for now.
     - name: gpu-operator
       type: Helm
+      manifestFiles:
+        # Ships a ConfigMap with options nvidia NVreg_GrdmaPciTopoCheckOverride=1.
+        # driver.kernelModuleConfig.name (below) points the ClusterPolicy at it,
+        # so the driver DaemonSet mounts nvidia.conf and the kernel loads with
+        # the flag set — required on GB200+EFA for EFA dma-buf attach to the
+        # Grace PCI topology. Without it, NCCL silently falls back to Socket.
+        - components/gpu-operator/manifests/kernel-module-params.yaml
       dependencyRefs:
         - nfd
         - cert-manager
@@ -46,6 +56,13 @@ spec:
           enabled: true
         gdrcopy:
           enabled: true
+        driver:
+          # 580.126.20 is NVIDIA's recommended floor for GB200+EFA; the global
+          # default (580.105.08 in components/gpu-operator/values.yaml) stays
+          # unchanged for H100/B200 and non-EKS GB200 recipes.
+          version: 580.126.20
+          kernelModuleConfig:
+            name: nvidia-kernel-module-params
 
     # GB200 uses nodewright no-op: the H100 tuning packages (nvidia-setup,
     # nvidia-tuned) are not compatible with GB200's ARM64 host CPU and
@@ -65,6 +82,20 @@ spec:
   # Validation checks for GB200 on EKS training workloads.
   # Defined at the intent layer (not OS-specific) so all OS variants inherit them.
   validation:
+    performance:
+      # NET exercises EFA (the external interconnect) and NVLS exercises MNNVL
+      # across the NVL72 IMEX domain — each variant asserts its transport
+      # actually carried traffic, so a silent fallback to Socket or NET does
+      # not masquerade as a pass. Thresholds are sized for a 2-node GB200 pair
+      # and will be raised once production NVL72 data is available.
+      checks:
+        - nccl-all-reduce-bw-net
+        - nccl-all-reduce-bw-nvls
+      constraints:
+        - name: nccl-all-reduce-bw-net
+          value: ">= 40"
+        - name: nccl-all-reduce-bw-nvls
+          value: ">= 500"
     conformance:
       checks:
         - platform-health