diff --git a/adafruit_neopxl8.py b/adafruit_neopxl8.py
index b7d0eb0..1ee5404 100644
--- a/adafruit_neopxl8.py
+++ b/adafruit_neopxl8.py
@@ -32,7 +32,7 @@
 bitloop:
     pull ifempty [1]     ; don't start outputting HIGH unless data is available (always-low part)
     mov pins, ~ null [3] ; always-high part
-    out pins, 8 [4]      ; variable part
+    {}                   ; variable part
     mov pins, null       ; always-low part (last cycle is the 'pull ifempty' after wrap)
 
     jmp y--, bitloop     ; always-low part
@@ -46,7 +46,6 @@
     jmp top
 """
 
-_ASSEMBLED = adafruit_pioasm.assemble(_PROGRAM)
 
 # Pixel color order constants
 RGB = "RGB"
@@ -110,7 +109,7 @@ def __init__(
         brightness=1.0,
         auto_write=True,
         pixel_order=None,
-    ):
+    ):  # pylint: disable=too-many-locals
         if n % num_strands:
             raise ValueError("Length must be a multiple of num_strands")
         if not pixel_order:
@@ -124,25 +123,48 @@ def __init__(
             n, brightness=brightness, byteorder=pixel_order, auto_write=auto_write
         )
 
-        data_len = bpp * n * 8 // num_strands
+        if num_strands == 1:
+            data_len = bpp * n
+            pack = ">L"
+            osr = False
+            loop_count = 8 * data_len
+        else:
+            data_len = bpp * n * 8 // num_strands
+            pack = "<L"
+            osr = True
+            loop_count = data_len
         padding_count = -data_len % 4
 
+        self._num_strands = num_strands
         self._data = bytearray(8 + data_len + padding_count)
         self._data32 = memoryview(self._data).cast("L")
-        self._transposed = memoryview(self._data)[4 : 4 + data_len]
-        self._data[:4] = struct.pack("<L", data_len - 1)
-        self._data[-4:] = struct.pack("<L", 3840 * 2)
+        self._pixels = memoryview(self._data)[4 : 4 + data_len]
+        self._data[:4] = struct.pack(pack, loop_count - 1)
+        self._data[-4:] = struct.pack(pack, 3840)
 
         self._num_strands = num_strands
 
+        if num_strands == 8:
+            variable_part = "out pins, 8 [4]      ; variable part"
+        elif num_strands == 1:
+            variable_part = "out pins, 1 [4]      ; variable part"
+        else:
+            variable_part = f"""
+                out pins, {num_strands} [3]       ; variable part
+                out x, {8-num_strands}            ; variable part
+            """
+
+        program = _PROGRAM.format(variable_part)
+        assembled = adafruit_pioasm.assemble(program)
+
         self._sm = rp2pio.StateMachine(
-            _ASSEMBLED,
+            assembled,
             frequency=800_000 * 16,
             first_out_pin=data0,
-            out_pin_count=8,
+            out_pin_count=num_strands,
             first_set_pin=data0,
             auto_pull=False,
-            out_shift_right=True,
+            out_shift_right=osr,
         )
 
     def deinit(self):
@@ -177,5 +199,9 @@ def num_strands(self):
     def _transmit(self, buffer):
         while self._sm.pending:
             pass
-        bitops.bit_transpose(buffer, self._transposed, self._num_strands)
-        self._sm.background_write(self._data32)
+        if self.num_strands == 1:
+            self._pixels[:] = buffer
+            self._sm.background_write(self._data32, swap=True)
+        else:
+            bitops.bit_transpose(buffer, self._pixels, self._num_strands)
+            self._sm.background_write(self._data32)
diff --git a/examples/neopxl8_animations.py b/examples/neopxl8_animations.py
index 06bc417..8b64712 100644
--- a/examples/neopxl8_animations.py
+++ b/examples/neopxl8_animations.py
@@ -2,17 +2,12 @@
 #
 # SPDX-License-Identifier: Unlicense
 
-# Set up a separate animation on each of 8 strips with NeoPxl8
-
 import board
-from adafruit_led_animation.animation.rainbow import Rainbow
+import rainbowio
+import adafruit_ticks
 from adafruit_led_animation.animation.comet import Comet
-from adafruit_led_animation.animation.rainbowcomet import RainbowComet
-from adafruit_led_animation.animation.rainbowchase import RainbowChase
-from adafruit_led_animation.animation.chase import Chase
 from adafruit_led_animation.group import AnimationGroup
 from adafruit_led_animation.helper import PixelMap
-from adafruit_led_animation import color
 from adafruit_neopxl8 import NeoPxl8
 
 # Customize for your strands here
@@ -28,30 +23,43 @@
     num_pixels,
     num_strands=num_strands,
     auto_write=False,
-    brightness=0.07,
+    brightness=0.50,
 )
 
 
-def strand(i):
+def strand(n):
     return PixelMap(
         pixels,
-        range(i * strand_length, (i + 1) * strand_length),
+        range(n * strand_length, (n + 1) * strand_length),
         individual_pixels=True,
     )
 
 
+# Create the 8 virtual strands
 strands = [strand(i) for i in range(num_strands)]
 
-animations = AnimationGroup(
-    Comet(strands[0], 0.5, color.CYAN),
-    Comet(strands[1], 0.4, color.AMBER),
-    RainbowComet(strands[2], 0.3),
-    RainbowComet(strands[3], 0.7),
-    Chase(strands[4], 0.05, size=2, spacing=3, color=color.PURPLE),
-    RainbowChase(strands[5], 0.05, size=2, spacing=3),
-    Rainbow(strands[6], 0.6),
-    Rainbow(strands[7], 0.23, step=21),
-)
+# For each strand, create a comet animation of a different color
+animations = [
+    Comet(strand, 0.02, rainbowio.colorwheel(3 * 32 * i), ring=True)
+    for i, strand in enumerate(strands)
+]
+
+# Advance the animations by varying amounts so that they become staggered
+for i, animation in enumerate(animations):
+    animation._tail_start = 30 * 5 * i // 8  # pylint: disable=protected-access
+
+# Group them so we can run them all at once
+animations = AnimationGroup(*animations)
 
+# Run the animations and report on the speed in frame per secodn
+t0 = adafruit_ticks.ticks_ms()
+frame_count = 0
 while True:
     animations.animate()
+    frame_count += 1
+    t1 = adafruit_ticks.ticks_ms()
+    dt = adafruit_ticks.ticks_diff(t1, t0)
+    if dt > 1000:
+        print(f"{frame_count * 1000/dt:.1f}fps")
+        t0 = t1
+        frame_count = 0