diff --git a/README.md b/README.md
index 63c1553..da9cc24 100644
--- a/README.md
+++ b/README.md
@@ -49,6 +49,7 @@ This path uses:
 Important:
 - `gpudemod_kernels.dll` must be present next to `sdrd.exe` or in `internal/demod/gpudemod/build/`
 - `build-sdrplay.ps1` prepares the runtime DLL placement and PATH setup for SDRplay + CUDA DLLs
+- GPU validation is disabled by default for performance; enable it with `SDR_GPU_VALIDATE=1` when debugging
 - older experimental Windows build scripts were removed to avoid confusion
 
 ### Linux
diff --git a/internal/demod/gpudemod/gpudemod_windows.go b/internal/demod/gpudemod/gpudemod_windows.go
index feda60d..28a67a4 100644
--- a/internal/demod/gpudemod/gpudemod_windows.go
+++ b/internal/demod/gpudemod/gpudemod_windows.go
@@ -435,8 +435,21 @@ func (e *Engine) Demod(iq []complex64, offsetHz float64, bw float64, mode DemodT
 		e.SetFIR(taps)
 	}
 	filtered, ok := e.tryCUDAFIR(shifted, len(taps))
-	e.lastFIRUsedGPU = ok && ValidateFIR(shifted, taps, filtered, 1e-3)
-	if !e.lastFIRUsedGPU {
+	if ok {
+		if validationEnabled() {
+			e.lastFIRUsedGPU = ValidateFIR(shifted, taps, filtered, 1e-3)
+			if !e.lastFIRUsedGPU {
+				ftaps := make([]float64, len(taps))
+				for i, v := range taps {
+					ftaps[i] = float64(v)
+				}
+				filtered = dsp.ApplyFIR(shifted, ftaps)
+			}
+		} else {
+			e.lastFIRUsedGPU = true
+		}
+	}
+	if filtered == nil {
 		ftaps := make([]float64, len(taps))
 		for i, v := range taps {
 			ftaps[i] = float64(v)
diff --git a/internal/demod/gpudemod/validation_runtime.go b/internal/demod/gpudemod/validation_runtime.go
new file mode 100644
index 0000000..40cc976
--- /dev/null
+++ b/internal/demod/gpudemod/validation_runtime.go
@@ -0,0 +1,9 @@
+//go:build cufft
+
+package gpudemod
+
+import "os"
+
+func validationEnabled() bool {
+	return os.Getenv("SDR_GPU_VALIDATE") == "1"
+}