APU: The NES Audio Engine

Built into the same chip as the CPU (2A03/2A07), the Audio Processing Unit gave the NES its distinctive sound. Two pulse waves handled melodies, a triangle wave provided bass, noise created percussion, and the delta modulation channel allowed rudimentary samples. Composers like Koji Kondo and Hip Tanaka turned these constraints into art.

Fast facts

Channels: 5 total (2 pulse, 1 triangle, 1 noise, 1 DMC).
Output: mono, mixed internally.
Sample rate: varies; DMC plays at 4.2–33.1 kHz depending on rate setting.
Integration: part of the 2A03 (NTSC) or 2A07 (PAL) CPU package.
NTSC/PAL differences: the 2A03 and 2A07 differ in CPU clock, frame-counter rate (~240 Hz vs ~200 Hz), DMC rate table, noise period table, and DMC DMA conflict timing — register semantics are otherwise identical.

Channel specifications

Pulse channels ($4000-$4007)

Feature	Details
Duty cycles	12.5%, 25%, 50%, and 25% negated — the hardware’s own label for the fourth
Frequency	~54 Hz to ~12.4 kHz (11-bit period; any period below 8 is silenced outright by the sweep unit)
Envelope	hardware decay or constant volume
Sweep	automatic pitch bend up/down
Length counter	automatic note cutoff

Four duty settings, three timbres. The fourth setting is not a 75% pulse in the hardware’s own terms — NESdev’s duty table calls it “25% negated”, and prints its output as 1 0 0 1 1 1 1 1 against 0 1 1 0 0 0 0 0 for plain 25%. It is the same shape with the high and low halves exchanged. Inverting a wave changes its phase, not its harmonic content, so 25% and 75% are indistinguishable played on their own. The difference only becomes audible against something else: another channel it can reinforce or cancel, or anything phase-sensitive downstream.

Ricoh’s specification for the chip lists all four because the hardware encodes four, not because a composer is offered four tones. In practice the palette is thin, bright (12.5%), reedy (25%, or its 75% twin) and hollow (50%) — which is a large part of why NES music is so recognisable. There was nowhere to hide.

Triangle channel ($4008-$400B)

Feature	Details
Waveform	32-step triangle (no volume control)
Frequency	up to f_CPU/32, about 55.9 kHz on NTSC — far above hearing (the timer is clocked every CPU cycle, where the other channels’ are clocked every second cycle)
Usage	bass lines, soft leads
Quirk	a timer of zero is how games silence the channel; the cost is a pop when an audible note resumes

Noise channel ($400C-$400F)

Feature	Details
Modes	long (32,767 steps) or short — 93 or 31 steps, depending where the shift register stood when the mode flag was set
Frequency	16 fixed rates
Usage	drums, explosions, sound effects

DMC channel ($4010-$4013)

Feature	Details
Format	1-bit delta-encoded samples
Sample rates	16 fixed rates (4.2-33.1 kHz)
Memory	samples must reside in $C000-$FFFF
Usage	voice, drums, complex sounds

Register summary

Address	Channel	Purpose
$4000-$4003	Pulse 1	Duty, envelope, sweep, period, length
$4004-$4007	Pulse 2	Same as Pulse 1
$4008-$400B	Triangle	Linear counter, period, length
$400C-$400F	Noise	Envelope, mode, period, length
$4010-$4013	DMC	Rate, sample address, length, output
$4015	All	Channel enable/status
$4017	Frame	Frame counter mode

Programming techniques

Arpeggios: rapidly cycle through notes to simulate chords.
Duty cycling: switch pulse duty mid-note for timbre variation — between 12.5%, 25% and 50%. Switching between 25% and 75% changes nothing audible on its own, for the reason given above.
Volume envelopes: software-controlled volume for more expressive decay.
Echo: delay audio data in software for pseudo-reverb.
DMC percussion: use samples for punchy drums unavailable in synthesis.

Expansion audio

Some cartridges added extra sound channels:

Mapper	Extra channels
VRC6	2 pulse + 1 sawtooth
VRC7	6 FM synthesis channels
Sunsoft 5B	3 PSG channels (AY-3-8910 compatible)
Namco 163	up to 8 wavetable channels
FDS	1 wavetable channel

Five channels of chiptune magic