diff --git a/transcode_bench.py b/transcode_bench.py
index 2d64e93..4eea50e 100755
--- a/transcode_bench.py
+++ b/transcode_bench.py
@@ -486,6 +486,7 @@ class Benchmark:
         self.hwaccel_args = hwaccel_args
         self.accel_name = accel_name
         self.hw_decode = hw_decode
+        self.hit_nvenc_limit = False  # Track if we hit NVIDIA session limit
 
         # Determine HEVC encoder (hardware or software fallback)
         self.hevc_encoder = self._detect_hevc_encoder()
@@ -578,9 +579,15 @@ class Benchmark:
             avg_fps = sum(job.fps for job in jobs) / len(jobs) if jobs else 0.0
             return True, avg_fps
 
-        # If failed, print first error for debugging
+        # If failed, check for NVENC session limit and print debug info
         for job in jobs:
             if not job.success and hasattr(job, 'stderr'):
+                # Check for NVIDIA session limit error
+                if 'OpenEncodeSessionEx failed' in job.stderr or 'incompatible client key' in job.stderr:
+                    self.hit_nvenc_limit = True
+                    print(f"✗ (NVIDIA concurrent session limit reached)")
+                    return False, 0.0
+
                 # Skip FFmpeg header and show actual error (last 10 non-empty lines)
                 error_lines = [line for line in job.stderr.split('\n') if line.strip()]
                 if len(error_lines) > 10:
@@ -644,8 +651,8 @@ class Benchmark:
                     print(f"✗ (failed)")
                 high = mid - 1
 
-        # If we have headroom, try one more stream
-        if max_streams > 0:
+        # If we have headroom, try one more stream (unless we hit NVENC limit)
+        if max_streams > 0 and not self.hit_nvenc_limit:
             print(f"Testing {max_streams + 1} simultaneous streams (checking headroom)...", end=' ', flush=True)
             success, avg_fps = self.run_parallel_transcodes(max_streams + 1)
 
@@ -655,7 +662,8 @@ class Benchmark:
             else:
                 if success:
                     print(f"✗ (avg {avg_fps:.1f} fps - below real-time)")
-                else:
+                elif not self.hit_nvenc_limit:
+                    # Only print "failed" if we didn't already print NVENC limit message
                     print(f"✗ (failed)")
 
         return max_streams
@@ -735,10 +743,21 @@ def main():
         print(f"Transcode Task: 1080p H.264 → 720p H.265")
         print(f"Maximum Simultaneous Streams: {max_streams}")
         print(f"(at {args.min_fps} FPS or better)")
+
+        # Show NVIDIA session limit warning if hit
+        if benchmark.hit_nvenc_limit:
+            print("\n⚠ IMPORTANT: Hit NVIDIA driver concurrent session limit!")
+            print("  Your GPU may be capable of more streams, but NVIDIA")
+            print("  GeForce drivers artificially limit concurrent NVENC sessions.")
+            print("  Typical limits: 2-5 sessions (older), 5-8 (RTX 3000/4000)")
+            print("  Quadro/Tesla cards have unlimited sessions.")
+
         print("=" * 60)
 
         # Also output just the number for easy parsing
         print(f"\nBenchmark Score: {max_streams}")
+        if benchmark.hit_nvenc_limit:
+            print("(Limited by NVIDIA driver, not hardware performance)")
 
     finally:
         # Cleanup