float16

WARNING

float16 is not implemented in the current release.

Property	Value
Size	2 bytes (16 bits)
Format	IEEE 754 binary16 (half precision)
Approximate range	±6.1 × 10-5 to ±6.5 × 104
Decimal precision	~3–4 digits
Literal suffix	f16
Hardware support	Native on GPUs and recent CPUs; emulated elsewhere

float16 is a half-precision float, fully IEEE 754-2019 compliant. Special values (Inf, -Inf, NaN) are supported, and division by zero produces Inf or NaN rather than a fatal error. Arithmetic may be emulated in software on CPU targets that lack native FP16 instructions.

Literals

let half: float16 = 0.0625;  // explicit annotation
let v = 1.0f16;              // type suffix

Unsuffixed floating-point literals default to float64; request float16 with the f16 suffix or an annotation. Decimal and scientific forms (6.5e4) are both accepted — see Literals.

Typical Use Cases

• GPU compute kernels and graphics pipelines
• Machine-learning inference

Arithmetic

Operator	Description	Compound
+	Addition	+=
-	Subtraction	-=
*	Multiplication	*=
/	Division	/=
%	Remainder	%=
**	Exponentiation	n/a

let a: float16 = 10.0;
let b: float16 = 3.0;

let sum  = a + b;   // 13.0
let quot = a / b;   // 3.33…  (rounded to the nearest representable value)
let rem  = a % b;   // 1.0

Both operands must share the same type; mixed-width expressions require an explicit cast. Arithmetic rounds to nearest (ties to even), unary negation (-x) flips the sign, and results that exceed the range overflow to Inf. Subtracting nearly equal values loses significant digits (catastrophic cancellation) — reorder the computation or use a wider type when accuracy matters.

Division by zero never fails: it yields signed infinity, and 0.0 / 0.0 yields NaN. Inf and NaN then propagate — Inf - Inf and Inf * 0.0 are both NaN.

let p =  1.0 / 0.0;  //  Inf
let n = -1.0 / 0.0;  // -Inf
let u =  0.0 / 0.0;  //  NaN

Comparison

Operator	Description	Result
==	Equal	bool
!=	Not equal	bool
<	Less than	bool
<=	Less than or equal	bool
>	Greater than	bool
>=	Greater than or equal	bool

Both operands must have the same type; comparing a different float width, or a float with an integer, is a compile-time error — cast explicitly first.

NaN is unordered: every comparison involving it returns false, including NaN == NaN. Detect it with IsNan, never ==. For the same reason, avoid == between computed results and compare within a tolerance instead:

let x = 0.1 + 0.2;
let close = Abs(x - 0.3) < 1.0e-3;  // not  x == 0.3

Conversion

Rux implicitly widens to any larger float type, so no precision is lost. Narrowing must be explicit and rounds to nearest (ties to even).

let x: float16 = 1.5;
let wide: float32 = x;       // widening — implicit
let narrow = x as float8;    // narrowing — explicit cast

Float-to-integer conversions are always explicit and truncate toward zero; casting NaN or Inf to an integer raises a fatal error in debug builds.

let whole = x as int32;      // 1   truncates toward zero

See Also

• float64 — the default double-precision float
• float32 — the next precision up