diff --git a/platform/plat_compat.h b/platform/plat_compat.h
index 551e100..0b46caa 100644
--- a/platform/plat_compat.h
+++ b/platform/plat_compat.h
@@ -55,7 +55,7 @@
 
     // platform name for the version string
     #define GEX_PLATFORM "STM32F103-Bluepill"
-    #define PLAT_AHB_CLOCK 72e6
+    #define PLAT_AHB_MHZ 72
 
     // feature flags
     #define PLAT_FLASHBANKS 1
@@ -106,7 +106,7 @@
 
     // platform name for the version string
     #define GEX_PLATFORM "STM32F072-Discovery"
-    #define PLAT_AHB_CLOCK 48e6
+    #define PLAT_AHB_MHZ 48
 
     #include <stm32f0xx.h>
     #include <stm32f0xx_ll_adc.h>
@@ -153,7 +153,7 @@
 
     // platform name for the version string
     #define GEX_PLATFORM "STM32F303-Discovery"
-    #define PLAT_AHB_CLOCK 72e6
+    #define PLAT_AHB_MHZ 72
 
     #include <stm32f3xx.h>
     #include <stm32f3xx_hal.h>
@@ -203,7 +203,7 @@
 
     // platform name for the version string
     #define GEX_PLATFORM "STM32F407-Discovery"
-    #define PLAT_AHB_CLOCK 168e6
+    #define PLAT_AHB_MHZ 168
 
     #define PLAT_USB_PHYCLOCK 1
     #define PLAT_USB_OTGFS 1
diff --git a/units/neopixel/ws2812.c b/units/neopixel/ws2812.c
index 9ecea62..00981ba 100644
--- a/units/neopixel/ws2812.c
+++ b/units/neopixel/ws2812.c
@@ -1,10 +1,12 @@
 #include "platform.h"
 #include "ws2812.h"
 
-#define PLAT_NEOPIXEL_SHORT (uint32_t)(800e-9*PLAT_AHB_CLOCK)
-#define PLAT_NEOPIXEL_LONG (uint32_t)(400e-9*PLAT_AHB_CLOCK)
+#define FREQ_STEP (PLAT_AHB_MHZ/20.0f)
+#define NPX_DELAY_SHORT (uint32_t)(FREQ_STEP*1.5f)
+#define NPX_DELAY_LONG  (uint32_t)(FREQ_STEP*3.5f)
+#define NPX_DELAY_SHOW  (uint32_t)(FREQ_STEP*60)
 
-static //inline __attribute__((always_inline))
+static inline __attribute__((always_inline))
 void ws2812_byte(GPIO_TypeDef *port, uint32_t ll_pin, uint8_t b)
 {
 	for (register volatile uint8_t i = 0; i < 8; i++) {
@@ -12,17 +14,13 @@ void ws2812_byte(GPIO_TypeDef *port, uint32_t ll_pin, uint8_t b)
 
 		// duty cycle determines bit value
 		if (b & 0x80) {
-            __delay_cycles(PLAT_NEOPIXEL_LONG);
-            //for(uint32_t _i = 0; _i < 10; _i++) asm volatile("nop");
+            __asm_loop(NPX_DELAY_LONG);
             LL_GPIO_ResetOutputPin(port, ll_pin);
-            //for(uint32_t _i = 0; _i < 10; _i++) asm volatile("nop");
-            __delay_cycles(PLAT_NEOPIXEL_SHORT);
+            __asm_loop(NPX_DELAY_SHORT);
 		} else {
-            __delay_cycles(PLAT_NEOPIXEL_SHORT);
-            //for(uint32_t _i = 0; _i < 5; _i++) asm volatile("nop");
+            __asm_loop(NPX_DELAY_SHORT);
             LL_GPIO_ResetOutputPin(port, ll_pin);
-            __delay_cycles(PLAT_NEOPIXEL_LONG);
-            //for(uint32_t _i = 0; _i < 10; _i++) asm volatile("nop");
+            __asm_loop(NPX_DELAY_LONG);
 		}
 
 		b <<= 1; // shift to next bit
@@ -43,7 +41,11 @@ void ws2812_load_raw(GPIO_TypeDef *port, uint32_t ll_pin, uint8_t *rgbs, uint32_
 		ws2812_byte(port, ll_pin, b);
 	}
     vPortExitCritical();
-	// TODO: Delay 50 us
+    __asm_loop(NPX_DELAY_SHOW);
+
+    LL_GPIO_SetOutputPin(port, ll_pin);
+    __asm_loop(NPX_DELAY_SHORT);
+    LL_GPIO_ResetOutputPin(port, ll_pin);
 }
 
 /** Set many RGBs from uint32 stream */
@@ -69,7 +71,11 @@ void ws2812_load_sparse(GPIO_TypeDef *port, uint32_t ll_pin, uint8_t *rgbs, uint
 		ws2812_byte(port, ll_pin, b);
 	}
     vPortExitCritical();
-	// TODO: Delay 50 us
+    __asm_loop(NPX_DELAY_SHOW);
+
+    LL_GPIO_SetOutputPin(port, ll_pin);
+    __asm_loop(NPX_DELAY_SHORT);
+    LL_GPIO_ResetOutputPin(port, ll_pin);
 }
 
 /** Set many RGBs */
@@ -80,5 +86,9 @@ void ws2812_clear(GPIO_TypeDef *port, uint32_t ll_pin, uint32_t count)
 		ws2812_byte(port, ll_pin, 0);
 	}
     vPortExitCritical();
-	// TODO: Delay 50 us
+    __asm_loop(NPX_DELAY_SHOW);
+
+    LL_GPIO_SetOutputPin(port, ll_pin);
+    __asm_loop(NPX_DELAY_SHORT);
+    LL_GPIO_ResetOutputPin(port, ll_pin);
 }
diff --git a/utils/cortex_utils.h b/utils/cortex_utils.h
index d7ed7cd..a344e7d 100644
--- a/utils/cortex_utils.h
+++ b/utils/cortex_utils.h
@@ -23,10 +23,18 @@ static inline bool isDynAlloc(const void *obj)
            && ((uint32_t)obj < (uint32_t)__SP);
 }
 
-static inline void __delay_cycles(uint32_t cycles)
-{
-    uint32_t l = cycles/3;
-    asm volatile( "0:" "sub %[count], #1;" "bne 0b;" :[count]"+r"(l) );
-}
+/** Tight asm loop */
+#define __asm_loop(cycles) \
+do { \
+    register uint32_t _count = cycles; \
+    asm volatile( \
+        ".syntax unified\n" \
+        ".thumb\n" \
+            "ldr r4, =%0\n" \
+        "0:" \
+            "subs r4, #1\n" \
+            "bne 0b\n" \
+        : /* no outputs */ : "g" (cycles) : "r4"); \
+} while(0)
 
 #endif //GEX_CORTEX_UTILS_H