diff --git a/tests/hash/CMakeLists.txt b/tests/hash/CMakeLists.txt
index 995bb090d..ce3757b4e 100644
--- a/tests/hash/CMakeLists.txt
+++ b/tests/hash/CMakeLists.txt
@@ -62,7 +62,8 @@ if (WBAG_CUDA)
         CUDA_STANDARD 17
         CUDA_STANDARD_REQUIRED YES
         POSITION_INDEPENDENT_CODE ON)
-      target_link_libraries(wbag_gpu_bench PRIVATE CUDA::cudart common cncrypto cryptonote_basic)
+      target_link_libraries(wbag_gpu_bench PRIVATE CUDA::cudart CUDA::cuda_driver common cncrypto cryptonote_basic)
+
 
       target_compile_definitions(hash-tests PRIVATE WBAG_HAVE_WBAG_GPU_BENCH=1)
       target_link_libraries(hash-tests PRIVATE wbag_gpu_bench)
diff --git a/tests/hash/wbag_gpu_bench.cu b/tests/hash/wbag_gpu_bench.cu
index aa6f105fd..d77baa671 100644
--- a/tests/hash/wbag_gpu_bench.cu
+++ b/tests/hash/wbag_gpu_bench.cu
@@ -1,4 +1,5 @@
 #include <cuda_runtime.h>
+#include <cuda.h>
 #include <cstdint>
 #include <cstdio>
 #include <vector>
@@ -130,6 +131,14 @@ static inline int parse_resident_mode_env() {
 #endif
 }
 
+static inline int cu_fail(CUresult r, const char* where){
+  if (r == CUDA_SUCCESS) return 0;
+  const char* name = nullptr; const char* desc = nullptr;
+  cuGetErrorName(r, &name); cuGetErrorString(r, &desc);
+  std::fprintf(stderr, "CUDA driver %s: %s (%s)\n", where, name ? name : "err", desc ? desc : "");
+  return 1;
+}
+
 int wbag_gpu_bench_once(const cryptonote::wbag::CacheView& view, int threads, int trials, double& hashes_per_sec_out)
 {
   if (threads <= 0 || trials <= 0) return 1;
@@ -137,14 +146,16 @@ int wbag_gpu_bench_once(const cryptonote::wbag::CacheView& view, int threads, in
   int devs = 0;
   cudaError_t st = cudaGetDeviceCount(&devs);
   if (st != cudaSuccess || devs <= 0) {
-    std::fprintf(stderr, "CUDA get device count: %s\n", cudaGetErrorString(st));
-    return 100;
+    if (cu_fail(cuInit(0), "cuInit")) return 100;
+    int dcount = 0;
+    if (cu_fail(cuDeviceGetCount(&dcount), "cuDeviceGetCount")) return 100;
+    if (dcount <= 0) { std::fprintf(stderr, "CUDA driver reports 0 devices\n"); return 100; }
+    CUdevice dev; if (cu_fail(cuDeviceGet(&dev, 0), "cuDeviceGet(0)")) return 101;
+    CUcontext ctx; if (cu_fail(cuDevicePrimaryCtxRetain(&ctx, dev), "cuDevicePrimaryCtxRetain")) return 101;
+    if (cu_fail(cuCtxSetCurrent(ctx), "cuCtxSetCurrent")) return 101;
   }
   st = cudaSetDevice(0);
-  if (st != cudaSuccess) {
-    std::fprintf(stderr, "CUDA set device: %s\n", cudaGetErrorString(st));
-    return 101;
-  }
+  if (st != cudaSuccess) { std::fprintf(stderr, "CUDA set device: %s\n", cudaGetErrorString(st)); return 101; }
 
   const size_t entry_bytes     = WBAG_entry_bytes();
   const size_t cache_bytes     = view.size();
@@ -160,16 +171,9 @@ int wbag_gpu_bench_once(const cryptonote::wbag::CacheView& view, int threads, in
   if (!d_cache || d_cache_size != cache_bytes) {
     if (d_cache) { cudaFree(d_cache); d_cache = nullptr; d_cache_size = 0; }
     st = cudaMalloc(&d_cache, cache_bytes);
-    if (st != cudaSuccess) {
-      std::fprintf(stderr, "GPU: cache alloc failed (%zu bytes): %s\n", cache_bytes, cudaGetErrorString(st));
-      return 2;
-    }
+    if (st != cudaSuccess) { std::fprintf(stderr, "GPU: cache alloc failed (%zu bytes): %s\n", cache_bytes, cudaGetErrorString(st)); return 2; }
     st = cudaMemcpy(d_cache, view.data(), cache_bytes, cudaMemcpyHostToDevice);
-    if (st != cudaSuccess) {
-      std::fprintf(stderr, "GPU: cache memcpy failed: %s\n", cudaGetErrorString(st));
-      cudaFree(d_cache); d_cache = nullptr; d_cache_size = 0;
-      return 3;
-    }
+    if (st != cudaSuccess) { std::fprintf(stderr, "GPU: cache memcpy failed: %s\n", cudaGetErrorString(st)); cudaFree(d_cache); d_cache = nullptr; d_cache_size = 0; return 3; }
     d_cache_size = cache_bytes;
   }
 
@@ -178,10 +182,7 @@ int wbag_gpu_bench_once(const cryptonote::wbag::CacheView& view, int threads, in
 
   size_t free_b = 0, total_b = 0;
   st = cudaMemGetInfo(&free_b, &total_b);
-  if (st != cudaSuccess) {
-    std::fprintf(stderr, "GPU: mem info failed: %s\n", cudaGetErrorString(st));
-    return 4;
-  }
+  if (st != cudaSuccess) { std::fprintf(stderr, "GPU: mem info failed: %s\n", cudaGetErrorString(st)); return 4; }
   if (need_resident + (64ULL << 20) > free_b) {
     std::fprintf(stderr, "GPU: not enough VRAM for resident (%zu MiB needed, %zu MiB free). Set WBAG_GPU_RESIDENT=shared or reduce --threads.\n", need_resident >> 20, free_b >> 20);
     return 5;
@@ -189,10 +190,7 @@ int wbag_gpu_bench_once(const cryptonote::wbag::CacheView& view, int threads, in
 
   uint8_t* d_resident = nullptr;
   st = cudaMalloc(&d_resident, need_resident);
-  if (st != cudaSuccess) {
-    std::fprintf(stderr, "GPU: resident alloc failed (%zu bytes): %s\n", need_resident, cudaGetErrorString(st));
-    return 6;
-  }
+  if (st != cudaSuccess) { std::fprintf(stderr, "GPU: resident alloc failed (%zu bytes): %s\n", need_resident, cudaGetErrorString(st)); return 6; }
 
   const uint64_t base_seed = WBAG_resident_seed_from_cache(view.data(), view.size());