feat(gpu): remove dead NCv1/K80 R470 driver support

parts/linux/cloud-init/artifacts/acl/cse_install_acl.sh

@@ -175,8 +175,9 @@ installGPUDriverSysext() {
     #
     # NVIDIA_GPU_DRIVER_TYPE is set by AgentBaker based on the GPU SKU maps in
     # gpu_components.go. Converged sizes get "grid"; RTX PRO 6000 BSE v6 gets
-    # "grid-v20" (Ubuntu-only, rejected below); modern CUDA SKUs get "cuda-lts" and legacy
-    # NCv1 gets "cuda". Only grid vs non-grid matters here, so both take the CUDA path below.
+    # "grid-v20" (Ubuntu-only, rejected below); all other CUDA SKUs (including legacy
+    # NCv1/K80) get "cuda-lts". Only grid vs non-grid matters here, so the CUDA path
+    # below handles every non-grid SKU.
     # Legacy GPUs (T4, V100) require proprietary CUDA drivers; A100+ use NVIDIA open drivers.
     local vm_sku
     vm_sku=$(get_compute_sku)

parts/linux/cloud-init/artifacts/mariner/cse_install_mariner.sh

@@ -115,8 +115,9 @@ downloadGPUDrivers() {
     #
     # NVIDIA_GPU_DRIVER_TYPE is set by AgentBaker based on the GPU SKU maps in
     # gpu_components.go. Converged sizes get "grid"; RTX PRO 6000 BSE v6 gets
-    # "grid-v20" (Ubuntu-only, rejected below); modern CUDA SKUs get "cuda-lts" and legacy
-    # NCv1 gets "cuda". Only grid vs non-grid matters here, so both take the CUDA path below.
+    # "grid-v20" (Ubuntu-only, rejected below); all other CUDA SKUs (including legacy
+    # NCv1/K80) get "cuda-lts". Only grid vs non-grid matters here, so the CUDA path
+    # below handles every non-grid SKU.
     # Legacy GPUs (T4, V100) require proprietary CUDA drivers; A100+ use NVIDIA open drivers.
     KERNEL_VERSION=$(uname -r | sed 's/-/./g')
     VM_SKU=$(get_compute_sku)

pkg/agent/baker.go

@@ -1494,17 +1494,9 @@ func GetGPUDriverVersion(size string) string {
 	if useGridDrivers(size) {
 		return datamodel.NvidiaGridDriverVersion
 	}
-	if isStandardNCv1(size) {
-		return datamodel.Nvidia470CudaDriverVersion
-	}
 	return datamodel.NvidiaCudaDriverVersion
 }
 
-func isStandardNCv1(size string) bool {
-	tmp := strings.ToLower(size)
-	return strings.HasPrefix(tmp, "standard_nc") && !strings.Contains(tmp, "_v")
-}
-
 func useGridDrivers(size string) bool {
 	return datamodel.ConvergedGPUDriverSizes[strings.ToLower(size)]
 }
@@ -1527,20 +1519,22 @@ func GetAKSGPUImageSHA(size string) string {
 
 // GetGPUDriverType maps a GPU VM size to the aks-gpu image variant used to install its driver.
 // The value becomes NVIDIA_GPU_DRIVER_TYPE at provision time, which selects the container image
-// mcr.microsoft.com/aks/aks-gpu-<type>. Modern CUDA compute SKUs (T4, V100, A100, H100, H200, ...)
+// mcr.microsoft.com/aks/aks-gpu-<type>. All CUDA compute SKUs (T4, V100, A100, H100, H200, ...)
 // use the R580 LTS image (aks-gpu-cuda-lts): it retains Volta/V100 support that the newer aks-gpu-cuda
 // R595 line drops, is supported through Aug 2028, and is the branch the VHD driver prebake is built
-// against. Legacy NCv1 (K80) keeps the separate "cuda" path with its pinned R470 driver.
+// against.
+//
+// Legacy NCv1 (K80, Kepler) is intentionally NOT special-cased. NVIDIA R470 was the last branch to
+// support Kepler data-center GPUs, and AKS never published an aks-gpu-cuda R470 image (that tag always
+// 404'd), so no managed driver can actually run a K80. NCv1 therefore falls through to the default
+// cuda-lts path like any other compute SKU. The hardware is EOL and effectively unused on AKS.
 func GetGPUDriverType(size string) string {
 	if useGridV20Drivers(size) {
 		return "grid-v20"
 	}
 	if useGridDrivers(size) {
 		return "grid"
 	}
-	if isStandardNCv1(size) {
-		return "cuda"
-	}
 	return "cuda-lts"
 }
 

pkg/agent/baker_test.go

@@ -939,8 +939,11 @@ var _ = Describe("Test normalizeResourceGroupNameForLabel", func() {
 })
 
 var _ = Describe("GetGPUDriverVersion", func() {
-	It("should use 470 with nc v1", func() {
-		Expect(GetGPUDriverVersion("standard_nc6")).To(Equal(datamodel.Nvidia470CudaDriverVersion))
+	// NCv1 (K80, Kepler) is EOL: R470 was the last branch to support it and AKS never published an
+	// R470 image, so it is no longer special-cased and falls through to the default CUDA (cuda-lts)
+	// version like any other compute SKU.
+	It("should fall through to cuda-lts version for legacy nc v1 (K80)", func() {
+		Expect(GetGPUDriverVersion("standard_nc6")).To(Equal(datamodel.NvidiaCudaDriverVersion))
 	})
 	It("should use cuda with nc v3", func() {
 		Expect(GetGPUDriverVersion("standard_nc6_v3")).To(Equal(datamodel.NvidiaCudaDriverVersion))
@@ -967,8 +970,9 @@ var _ = Describe("GetGPUDriverType", func() {
 	It("should use cuda-lts with nc v3", func() {
 		Expect(GetGPUDriverType("standard_nc6_v3")).To(Equal("cuda-lts"))
 	})
-	It("should keep cuda (legacy R470) with nc v1 (K80)", func() {
-		Expect(GetGPUDriverType("standard_nc6")).To(Equal("cuda"))
+	// NCv1 (K80, Kepler) is EOL and no longer special-cased; it falls through to cuda-lts.
+	It("should fall through to cuda-lts for legacy nc v1 (K80)", func() {
+		Expect(GetGPUDriverType("standard_nc6")).To(Equal("cuda-lts"))
 	})
 	It("should use grid with nv v5", func() {
 		Expect(GetGPUDriverType("standard_nv6ads_a10_v5")).To(Equal("grid"))

pkg/agent/datamodel/gpu_components.go

@@ -8,8 +8,6 @@ import (
 	"github.com/Azure/agentbaker/parts"
 )
 
-const Nvidia470CudaDriverVersion = "cuda-470.82.01"
-
 //nolint:gochecknoglobals
 var (
 	NvidiaCudaDriverVersion    string