diff --git a/pkg/kubelet/gpu/nvidia/helpers.go b/pkg/kubelet/gpu/nvidia/helpers.go
index 0abedbf4014..a6ed7c4f935 100644
--- a/pkg/kubelet/gpu/nvidia/helpers.go
+++ b/pkg/kubelet/gpu/nvidia/helpers.go
@@ -18,14 +18,16 @@ package nvidia
 
 import "k8s.io/apimachinery/pkg/util/sets"
 
+type containerToGPU map[string]sets.String
+
 // podGPUs represents a list of pod to GPU mappings.
 type podGPUs struct {
-	podGPUMapping map[string]sets.String
+	podGPUMapping map[string]containerToGPU
 }
 
 func newPodGPUs() *podGPUs {
 	return &podGPUs{
-		podGPUMapping: map[string]sets.String{},
+		podGPUMapping: make(map[string]containerToGPU),
 	}
 }
 func (pgpu *podGPUs) pods() sets.String {
@@ -36,12 +38,26 @@ func (pgpu *podGPUs) pods() sets.String {
 	return ret
 }
 
-func (pgpu *podGPUs) insert(podUID string, device string) {
+func (pgpu *podGPUs) insert(podUID, contName string, device string) {
 	if _, exists := pgpu.podGPUMapping[podUID]; !exists {
-		pgpu.podGPUMapping[podUID] = sets.NewString(device)
-	} else {
-		pgpu.podGPUMapping[podUID].Insert(device)
+		pgpu.podGPUMapping[podUID] = make(containerToGPU)
 	}
+	if _, exists := pgpu.podGPUMapping[podUID][contName]; !exists {
+		pgpu.podGPUMapping[podUID][contName] = sets.NewString()
+	}
+	pgpu.podGPUMapping[podUID][contName].Insert(device)
+}
+
+func (pgpu *podGPUs) getGPUs(podUID, contName string) sets.String {
+	containers, exists := pgpu.podGPUMapping[podUID]
+	if !exists {
+		return nil
+	}
+	devices, exists := containers[contName]
+	if !exists {
+		return nil
+	}
+	return devices
 }
 
 func (pgpu *podGPUs) delete(pods []string) {
@@ -52,8 +68,10 @@ func (pgpu *podGPUs) delete(pods []string) {
 
 func (pgpu *podGPUs) devices() sets.String {
 	ret := sets.NewString()
-	for _, devices := range pgpu.podGPUMapping {
-		ret = ret.Union(devices)
+	for _, containerToGPU := range pgpu.podGPUMapping {
+		for _, deviceSet := range containerToGPU {
+			ret = ret.Union(deviceSet)
+		}
 	}
 	return ret
 }
diff --git a/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager.go b/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager.go
index 43b0e3b32a0..7d7176603e4 100644
--- a/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager.go
+++ b/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager.go
@@ -152,6 +152,12 @@ func (ngm *nvidiaGPUManager) AllocateGPU(pod *v1.Pod, container *v1.Container) (
 			return nil, fmt.Errorf("Failed to allocate GPUs because of issues with updating GPUs in use: %v", err)
 		}
 	}
+	// Check if GPUs have already been allocated. If so return them right away.
+	// This can happen if a container restarts for example.
+	if devices := ngm.allocated.getGPUs(string(pod.UID), container.Name); devices != nil {
+		glog.V(2).Infof("Found pre-allocated GPUs for container %q in Pod %q: %v", container.Name, pod.UID, devices.List())
+		return append(devices.List(), ngm.defaultDevices...), nil
+	}
 	// Get GPU devices in use.
 	devicesInUse := ngm.allocated.devices()
 	glog.V(5).Infof("gpus in use: %v", devicesInUse.List())
@@ -164,7 +170,7 @@ func (ngm *nvidiaGPUManager) AllocateGPU(pod *v1.Pod, container *v1.Container) (
 	ret := available.UnsortedList()[:gpusNeeded]
 	for _, device := range ret {
 		// Update internal allocated GPU cache.
-		ngm.allocated.insert(string(pod.UID), device)
+		ngm.allocated.insert(string(pod.UID), container.Name, device)
 	}
 	// Add standard devices files that needs to be exposed.
 	ret = append(ret, ngm.defaultDevices...)
@@ -222,9 +228,13 @@ func (ngm *nvidiaGPUManager) gpusInUse() (*podGPUs, error) {
 	if err != nil {
 		return nil, err
 	}
+	type containerIdentifier struct {
+		id   string
+		name string
+	}
 	type podContainers struct {
-		uid          string
-		containerIDs sets.String
+		uid        string
+		containers []containerIdentifier
 	}
 	// List of containers to inspect.
 	podContainersToInspect := []podContainers{}
@@ -240,21 +250,23 @@ func (ngm *nvidiaGPUManager) gpusInUse() (*podGPUs, error) {
 		if containers.Len() == 0 {
 			continue
 		}
-		containerIDs := sets.NewString()
+		// TODO: If kubelet restarts right after allocating a GPU to a pod, the container might not have started yet and so container status might not be available yet.
+		// Use an internal checkpoint instead or try using the CRI if its checkpoint is reliable.
+		var containersToInspect []containerIdentifier
 		for _, container := range pod.Status.ContainerStatuses {
 			if containers.Has(container.Name) {
-				containerIDs.Insert(container.ContainerID)
+				containersToInspect = append(containersToInspect, containerIdentifier{container.ContainerID, container.Name})
 			}
 		}
 		// add the pod and its containers that need to be inspected.
-		podContainersToInspect = append(podContainersToInspect, podContainers{string(pod.UID), containerIDs})
+		podContainersToInspect = append(podContainersToInspect, podContainers{string(pod.UID), containersToInspect})
 	}
 	ret := newPodGPUs()
 	for _, podContainer := range podContainersToInspect {
-		for _, containerId := range podContainer.containerIDs.List() {
-			containerJSON, err := ngm.dockerClient.InspectContainer(containerId)
+		for _, containerIdentifier := range podContainer.containers {
+			containerJSON, err := ngm.dockerClient.InspectContainer(containerIdentifier.id)
 			if err != nil {
-				glog.V(3).Infof("Failed to inspect container %q in pod %q while attempting to reconcile nvidia gpus in use", containerId, podContainer.uid)
+				glog.V(3).Infof("Failed to inspect container %q in pod %q while attempting to reconcile nvidia gpus in use", containerIdentifier.id, podContainer.uid)
 				continue
 			}
 
@@ -266,7 +278,7 @@ func (ngm *nvidiaGPUManager) gpusInUse() (*podGPUs, error) {
 			for _, device := range devices {
 				if isValidPath(device.PathOnHost) {
 					glog.V(4).Infof("Nvidia GPU %q is in use by Docker Container: %q", device.PathOnHost, containerJSON.ID)
-					ret.insert(podContainer.uid, device.PathOnHost)
+					ret.insert(podContainer.uid, containerIdentifier.name, device.PathOnHost)
 				}
 			}
 		}
diff --git a/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager_test.go b/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager_test.go
index aea168ba568..e4565525e49 100644
--- a/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager_test.go
+++ b/pkg/kubelet/gpu/nvidia/nvidia_gpu_manager_test.go
@@ -36,8 +36,8 @@ func (tapl *testActivePodsLister) GetRunningPods() ([]*v1.Pod, error) {
 	return tapl.activePods, nil
 }
 
-func makeTestPod(numContainers int) *v1.Pod {
-	quantity := resource.NewQuantity(1, resource.DecimalSI)
+func makeTestPod(numContainers, gpusPerContainer int) *v1.Pod {
+	quantity := resource.NewQuantity(int64(gpusPerContainer), resource.DecimalSI)
 	resources := v1.ResourceRequirements{
 		Limits: v1.ResourceList{
 			v1.ResourceNvidiaGPU: *quantity,
@@ -53,6 +53,7 @@ func makeTestPod(numContainers int) *v1.Pod {
 	}
 	for ; numContainers > 0; numContainers-- {
 		pod.Spec.Containers = append(pod.Spec.Containers, v1.Container{
+			Name:      string(uuid.NewUUID()),
 			Resources: resources,
 		})
 	}
@@ -75,7 +76,7 @@ func TestMultiContainerPodGPUAllocation(t *testing.T) {
 	as.Equal(len(gpusInUse.devices()), 0)
 
 	// Allocated GPUs for a pod with two containers.
-	pod := makeTestPod(2)
+	pod := makeTestPod(2, 1)
 	// Allocate for the first container.
 	devices1, err := testGpuManager.AllocateGPU(pod, &pod.Spec.Containers[0])
 	as.Nil(err)
@@ -90,7 +91,7 @@ func TestMultiContainerPodGPUAllocation(t *testing.T) {
 	as.NotEqual(devices1, devices2, "expected containers to get different devices")
 
 	// further allocations should fail.
-	newPod := makeTestPod(2)
+	newPod := makeTestPod(2, 1)
 	devices1, err = testGpuManager.AllocateGPU(newPod, &newPod.Spec.Containers[0])
 	as.NotNil(err, "expected gpu allocation to fail. got: %v", devices1)
 
@@ -126,7 +127,7 @@ func TestMultiPodGPUAllocation(t *testing.T) {
 	as.Equal(len(gpusInUse.devices()), 0)
 
 	// Allocated GPUs for a pod with two containers.
-	podA := makeTestPod(1)
+	podA := makeTestPod(1, 1)
 	// Allocate for the first container.
 	devicesA, err := testGpuManager.AllocateGPU(podA, &podA.Spec.Containers[0])
 	as.Nil(err)
@@ -135,10 +136,48 @@ func TestMultiPodGPUAllocation(t *testing.T) {
 	podLister.activePods = append(podLister.activePods, podA)
 
 	// further allocations should fail.
-	podB := makeTestPod(1)
+	podB := makeTestPod(1, 1)
 	// Allocate for the first container.
 	devicesB, err := testGpuManager.AllocateGPU(podB, &podB.Spec.Containers[0])
 	as.Nil(err)
 	as.Equal(len(devicesB), 1)
 	as.NotEqual(devicesA, devicesB, "expected pods to get different devices")
 }
+
+func TestPodContainerRestart(t *testing.T) {
+	podLister := &testActivePodsLister{}
+
+	testGpuManager := &nvidiaGPUManager{
+		activePodsLister: podLister,
+		allGPUs:          sets.NewString("/dev/nvidia0", "/dev/nvidia1"),
+		allocated:        newPodGPUs(),
+		defaultDevices:   []string{"/dev/nvidia-smi"},
+	}
+
+	// Expect that no devices are in use.
+	gpusInUse, err := testGpuManager.gpusInUse()
+	as := assert.New(t)
+	as.Nil(err)
+	as.Equal(len(gpusInUse.devices()), 0)
+
+	// Make a pod with one containers that requests two GPUs.
+	podA := makeTestPod(1, 2)
+	// Allocate GPUs
+	devicesA, err := testGpuManager.AllocateGPU(podA, &podA.Spec.Containers[0])
+	as.Nil(err)
+	as.Equal(len(devicesA), 3)
+
+	podLister.activePods = append(podLister.activePods, podA)
+
+	// further allocations should fail.
+	podB := makeTestPod(1, 1)
+	_, err = testGpuManager.AllocateGPU(podB, &podB.Spec.Containers[0])
+	as.NotNil(err)
+
+	// Allcate GPU for existing Pod A.
+	// The same gpus must be returned.
+	devicesAretry, err := testGpuManager.AllocateGPU(podA, &podA.Spec.Containers[0])
+	as.Nil(err)
+	as.Equal(len(devicesA), 3)
+	as.True(sets.NewString(devicesA...).Equal(sets.NewString(devicesAretry...)))
+}
diff --git a/test/e2e_node/gpus.go b/test/e2e_node/gpus.go
index 73763237bb9..1e3eb73748c 100644
--- a/test/e2e_node/gpus.go
+++ b/test/e2e_node/gpus.go
@@ -18,6 +18,7 @@ package e2e_node
 
 import (
 	"fmt"
+	"time"
 
 	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
@@ -79,6 +80,20 @@ var _ = framework.KubeDescribe("GPU [Serial]", func() {
 			podSuccess := makePod(gpusAvailable.Value(), "gpus-success")
 			podSuccess = f.PodClient().CreateSync(podSuccess)
 
+			By("Checking the containers in the pod had restarted at-least twice successfully thereby ensuring GPUs are reused")
+			const minContainerRestartCount = 2
+			Eventually(func() bool {
+				p, err := f.ClientSet.Core().Pods(f.Namespace.Name).Get(podSuccess.Name, metav1.GetOptions{})
+				if err != nil {
+					framework.Logf("failed to get pod status: %v", err)
+					return false
+				}
+				if p.Status.ContainerStatuses[0].RestartCount < minContainerRestartCount {
+					return false
+				}
+				return true
+			}, time.Minute, time.Second).Should(BeTrue())
+
 			By("Checking if the pod outputted Success to its logs")
 			framework.ExpectNoError(f.PodClient().MatchContainerOutput(podSuccess.Name, podSuccess.Name, "Success"))
 
@@ -115,12 +130,13 @@ func makePod(gpus int64, name string) *v1.Pod {
 			v1.ResourceNvidiaGPU: *resource.NewQuantity(gpus, resource.DecimalSI),
 		},
 	}
-	gpuverificationCmd := fmt.Sprintf("if [[ %d -ne $(ls /dev/ | egrep '^nvidia[0-9]+$') ]]; then exit 1; fi; echo Success; sleep 10240 ", gpus)
+	gpuverificationCmd := fmt.Sprintf("if [[ %d -ne $(ls /dev/ | egrep '^nvidia[0-9]+$') ]]; then exit 1; fi; echo Success", gpus)
 	return &v1.Pod{
 		ObjectMeta: metav1.ObjectMeta{
 			Name: name,
 		},
 		Spec: v1.PodSpec{
+			RestartPolicy: v1.RestartPolicyAlways,
 			Containers: []v1.Container{
 				{
 					Image:     "gcr.io/google_containers/busybox:1.24",