diff --git a/runner/cmd/shim/main.go b/runner/cmd/shim/main.go
index 116e16b50c..50a163a0eb 100644
--- a/runner/cmd/shim/main.go
+++ b/runner/cmd/shim/main.go
@@ -269,7 +269,7 @@ func start(ctx context.Context, args shim.CLIArgs, serviceMode bool) (err error)
 	var dcgmExporter *dcgm.DCGMExporter
 	var dcgmWrapper dcgm.DCGMWrapperInterface
 
-	if gpu.GetGpuVendor() == gpu.GpuVendorNvidia {
+	if gpu.GetGpuVendor(ctx) == gpu.GpuVendorNvidia {
 		dcgmExporterPath, err := dcgm.GetDCGMExporterExecPath(ctx)
 		if err == nil {
 			interval := time.Duration(args.DCGMExporter.Interval * int(time.Millisecond))
diff --git a/runner/internal/common/gpu/gpu.go b/runner/internal/common/gpu/gpu.go
index 72ae83bb56..0a2c5afac8 100644
--- a/runner/internal/common/gpu/gpu.go
+++ b/runner/internal/common/gpu/gpu.go
@@ -1,8 +1,11 @@
 package gpu
 
 import (
+	"context"
 	"errors"
 	"os"
+
+	"github.com/dstackai/dstack/runner/internal/common/log"
 )
 
 type GpuVendor string
@@ -15,26 +18,65 @@ const (
 	GpuVendorTenstorrent GpuVendor = "tenstorrent"
 )
 
-func GetGpuVendor() GpuVendor {
-	// FIXME: There might be errors other than os.ErrNotExist that are ignored silently.
-	// Propagate and log.
-	if _, err := os.Stat("/dev/kfd"); !errors.Is(err, os.ErrNotExist) {
-		return GpuVendorAmd
-	}
-	if _, err := os.Stat("/dev/nvidiactl"); !errors.Is(err, os.ErrNotExist) {
+func GetGpuVendor(ctx context.Context) GpuVendor {
+	// Some devices can be detected unambiguously -- they have unique device paths.
+	//
+	// The order within this group does not matter -- false positives are unlikely,
+	// but more likely options should be checked first (e.g., NVIDIA before Tenstorrent).
+	//
+	// This group **must** stay before the ambiguous one below, otherwise a host with
+	// an AMD iGPU and an NVIDIA dGPU (a common combination) is detected as AMD,
+	// see: https://github.com/dstackai/dstack/issues/4085
+
+	// NVIDIA
+	if checkPath(ctx, "/dev/nvidiactl") {
 		return GpuVendorNvidia
 	}
-	if _, err := os.Stat("/dev/accel"); !errors.Is(err, os.ErrNotExist) {
-		return GpuVendorIntel
+	// NVIDIA on WSL2
+	if checkPath(ctx, "/dev/dxg") && checkPath(ctx, "/usr/lib/wsl/lib/nvidia-smi") {
+		return GpuVendorNvidia
 	}
-	if _, err := os.Stat("/dev/tenstorrent"); !errors.Is(err, os.ErrNotExist) {
+	// Tenstorrent
+	if checkPath(ctx, "/dev/tenstorrent") {
 		return GpuVendorTenstorrent
 	}
-	if _, err := os.Stat("/dev/dxg"); !errors.Is(err, os.ErrNotExist) {
-		// WSL2
-		if _, err := os.Stat("/usr/lib/wsl/lib/nvidia-smi"); !errors.Is(err, os.ErrNotExist) {
-			return GpuVendorNvidia
-		}
+
+	// The following devices are tricky -- the same paths are used for devices that we
+	// support and expect and devices that we don't support and don't want to support, such
+	// as AMD iGPUs or AMD/Intel NPUs integrated into CPUs.
+	//
+	// The order **does** matter, since both paths can be present on the same host, e.g., an
+	// AMD dGPU (/dev/kfd) on a host with an AMD NPU (/dev/accel). We decided to check for AMD
+	// first and Intel Gaudi last, because /dev/kfd is the more reliable signal of the two:
+	// * /dev/kfd is at least vendor-specific -- it identifies the vendor but not the device class
+	// * /dev/accel is standardized[1] vendor-agnostic path used by (including but not limited to):
+	//   Intel (Habana Labs devices and NPU), AMD, Qualcomm, Rockchip -- it identifies neither,
+	//   so mapping it to Intel Gaudi is a last-resort guess
+	//
+	// Apparently, there are also far more legit AMD accelerators (e.g., Instinct) deployed in
+	// the wild than Intel Gaudi accelerators, so the guess we make more often is the better one.
+	//
+	// [1]: https://github.com/torvalds/linux/blob/master/Documentation/accel/introduction.rst
+
+	// AMD, **including** iGPU
+	if checkPath(ctx, "/dev/kfd") {
+		return GpuVendorAmd
+	}
+	// Intel/Habana Labs Gaudi OR some other compute accelerator
+	if checkPath(ctx, "/dev/accel") {
+		return GpuVendorIntel
 	}
+
 	return GpuVendorNone
 }
+
+func checkPath(ctx context.Context, path string) bool {
+	_, err := os.Stat(path)
+	if err == nil {
+		return true
+	}
+	if !errors.Is(err, os.ErrNotExist) {
+		log.Error(ctx, "Failed to check path while detecting accelerator", "path", path, "err", err)
+	}
+	return false
+}
diff --git a/runner/internal/runner/metrics/metrics.go b/runner/internal/runner/metrics/metrics.go
index 56c27a2bb1..34f22c731a 100644
--- a/runner/internal/runner/metrics/metrics.go
+++ b/runner/internal/runner/metrics/metrics.go
@@ -29,7 +29,7 @@ func NewMetricsCollector(ctx context.Context) (*MetricsCollector, error) {
 	if err != nil {
 		return nil, fmt.Errorf("get cgroup mount point: %w", err)
 	}
-	gpuVendor := gpu.GetGpuVendor()
+	gpuVendor := gpu.GetGpuVendor(ctx)
 	return &MetricsCollector{
 		cgroupMountPoint: cgroupMountPoint,
 		gpuVendor:        gpuVendor,
diff --git a/runner/internal/shim/host/gpu.go b/runner/internal/shim/host/gpu.go
index a9e7f97058..0526123da3 100644
--- a/runner/internal/shim/host/gpu.go
+++ b/runner/internal/shim/host/gpu.go
@@ -49,7 +49,7 @@ type GpuInfo struct {
 }
 
 func GetGpuInfo(ctx context.Context) []GpuInfo {
-	switch gpuVendor := gpu.GetGpuVendor(); gpuVendor {
+	switch gpuVendor := gpu.GetGpuVendor(ctx); gpuVendor {
 	case gpu.GpuVendorNvidia:
 		return getNvidiaGpuInfo(ctx)
 	case gpu.GpuVendorAmd: