diff --git a/src/dsp/helpers.rs b/src/dsp/helpers.rs
index 1625196..2d5ac15 100644
--- a/src/dsp/helpers.rs
+++ b/src/dsp/helpers.rs
@@ -561,8 +561,8 @@ impl DelayBuffer {
 
     #[inline]
     pub fn feed(&mut self, input: f32) {
-        self.wr = (self.wr + 1) % self.data.len();
         self.data[self.wr] = input;
+        self.wr = (self.wr + 1) % self.data.len();
     }
 
     #[inline]
diff --git a/src/dsp/node_delay.rs b/src/dsp/node_delay.rs
index 4c9bdde..a73e2d6 100644
--- a/src/dsp/node_delay.rs
+++ b/src/dsp/node_delay.rs
@@ -21,7 +21,6 @@ macro_rules! fa_delay_mode { ($formatter: expr, $v: expr, $denorm_v: expr) => {
 #[derive(Debug, Clone)]
 pub struct Delay {
     buffer:             Box<DelayBuffer>,
-    fb_sample:          f32,
     clock:              TriggerSampleClock,
 }
 
@@ -29,7 +28,6 @@ impl Delay {
     pub fn new(_nid: &NodeId) -> Self {
         Self {
             buffer:            Box::new(DelayBuffer::new()),
-            fb_sample:         0.0,
             clock:             TriggerSampleClock::new(),
         }
     }
@@ -108,43 +106,36 @@ impl DspNode for Delay {
         let mix  = inp::Delay::mix(inputs);
         let out  = out::Delay::sig(outputs);
 
-        let mut fb_s = self.fb_sample;
-
         if mode.i() == 0 {
             for frame in 0..ctx.nframes() {
                 let dry = inp.read(frame);
-                buffer.feed(dry + fb_s * denorm::Delay::fb(fb, frame));
 
                 let out_sample =
                     buffer.cubic_interpolate_at(
                         denorm::Delay::time(time, frame));
 
+                buffer.feed(dry + out_sample * denorm::Delay::fb(fb, frame));
+
                 out.write(frame,
                     crossfade(dry, out_sample,
                         denorm::Delay::mix(mix, frame).clamp(0.0, 1.0)));
-
-                fb_s = out_sample;
             }
         } else {
             for frame in 0..ctx.nframes() {
                 let dry = inp.read(frame);
-                buffer.feed(dry + fb_s * denorm::Delay::fb(fb, frame));
 
                 let clock_samples =
                     self.clock.next(denorm::Delay::trig(trig, frame));
-
                 let out_sample = buffer.at(clock_samples as usize);
 
+                buffer.feed(dry + out_sample * denorm::Delay::fb(fb, frame));
+
                 out.write(frame,
                     crossfade(dry, out_sample,
                         denorm::Delay::mix(mix, frame).clamp(0.0, 1.0)));
-
-                fb_s = out_sample;
             }
         }
 
-        self.fb_sample = fb_s;
-
         let last_frame = ctx.nframes() - 1;
         ctx_vals[0].set(out.read(last_frame));
     }
diff --git a/tests/node_delay.rs b/tests/node_delay.rs
index d81b7aa..b4ae726 100644
--- a/tests/node_delay.rs
+++ b/tests/node_delay.rs
@@ -225,7 +225,7 @@ fn check_node_delay_fb() {
     let idxs_big = collect_signal_changes(&res.0[..], 50);
 
     // We expect the signal to be delayed by 20ms:
-    assert_eq!(idxs_big, vec![(220, 106), (441, 53)]);
+    assert_eq!(idxs_big, vec![(220, 106), (440, 53)]);
 }
 
 #[test]
@@ -258,7 +258,7 @@ fn check_node_delay_fb_neg() {
 
     let idxs_big = collect_signal_changes(&res.0[..], 70);
 
-    assert_eq!(idxs_big, vec![(441, 100), (883, -100), (1325, 100)]);
+    assert_eq!(idxs_big, vec![(441, 100), (882, -100), (1323, 100)]);
 }
 
 
@@ -288,9 +288,19 @@ fn check_node_delay_fb_pos() {
     // Emit the trigger signal:
     pset_n(&mut matrix, test, "trig", 1.0);
 
-    let res = run_for_ms(&mut node_exec, 40.0);
+    let res = run_for_ms(&mut node_exec, 100.0);
 
     let idxs_big = collect_signal_changes(&res.0[..], 70);
 
-    assert_eq!(idxs_big, vec![(441, 100), (883, 100), (1325, 100)]);
+    assert_eq!(idxs_big, vec![
+        (441,           100),
+        (441 + 1 * 441, 100),
+        (441 + 2 * 441, 100),
+        (441 + 3 * 441, 100),
+        (441 + 4 * 441, 100),
+        (441 + 5 * 441, 100),
+        (441 + 6 * 441, 100),
+        (441 + 7 * 441, 100),
+        (441 + 8 * 441, 100),
+    ]);
 }