IEEE 754 Precision Reference — the Formats’ Machine-Precision States

The machine-precision states of the IEEE 754 binary formats as the books tabulate them: Book Thirteen’s first chapter tabulates the four formats — FP16, FP32, FP64, FP128 — each at its machine epsilon ε (the spacing of representable numbers at 1) and its exponent range, and Book Seventeen’s quantization chapter carries the identity’s second table, the step Δq = 1/(2N − 1) at five depths. This page returns the format’s ε and range at the stated format — a reference return: the tables, citable — and computes each table’s governing relation for the visitor’s own stated precision or depth. The tables stand below as the page’s reference panels, in the books’ own printed states.

Machine Epsilon — ε = 2−p

The relation as the book states it: ε = 2−p with p the format’s significand precision, in the book’s stated convention — ε at the 1.0 spacing state. Choose a row of the table, or state a precision of your own. The page opens on the table’s FP64 row.

—

Reference Panel — Book Thirteen’s Table (Ch 1)

FP16 — ε = 2−11 ≈ 4.88 × 10−4; range 6.10 × 10−5 to 6.55 × 104, as the table states them.

FP32 — ε = 2−24 ≈ 5.96 × 10−8; range 1.18 × 10−38 to 3.40 × 1038.

FP64 — ε = 2−53 ≈ 1.11 × 10−16; range 2.23 × 10−308 to 1.79 × 10308.

FP128 — ε = 2−113 ≈ 9.63 × 10−35; range 3.36 × 10−4932 to 1.18 × 104932.

The rows’ standing — every sampled row of the table recomputes exactly at its stated precision (the census’s completeness sample, Ch 1): the ε column by the relation ε = 2−p, the ranges as the smallest normal to the largest finite of each format’s exponent field. Beside the table, the same chapter’s prose restates the FP32 and FP64 epsilons in a piecewise statement and records the subnormal threshold 2−1022 — recorded beside the table, not a row of it.

Reference Panel — Book Seventeen’s Quantization Table (Ch 2 §2.6.2)

8-Bit Integer — Δq printed 3.9216 × 10−3 — the relation’s value, 3.92157 × 10−3, rounded at the printed precision · ΔE00 ∼3.20.

10-Bit Integer — Δq printed 9.7751 × 10−4, standing in the truncation state at the final digit — the relation returns 9.77517 × 10−4 · ΔE00 ∼0.80.

12-Bit Integer — Δq printed 2.4420 × 10−4 — the relation returns 2.44200 × 10−4 · ΔE00 ∼0.20.

16-Bit IEEE Half-Float — one of this table’s two counted IEEE standings (the identity recurs at Ch 22’s payload diagram, “IEEE 754 Float16”) · ΔE00 ∼0.05.

32-Bit IEEE Float — the second counted IEEE standing (Ch 29’s protocol frame table carries three “IEEE 754 Float” field types) · ΔE00 <0.001.

The ΔE00 column stands as the table’s own stated standings: no model is stated from which to recompute them, and none is constructed on this page.

The identity across the corpus — beyond its two tables, the 754 identity stands at specification and furniture extent through the books: Fifteen’s and Sixteen’s specification standings; Eighteen’s representation standing (an “IEEE754 representation” field in a register table); the furniture standings of Twenty, Twenty-Two through Twenty-Eight, and Thirty-Two; the float-format specifications of Thirty-Three (two seats), Thirty-Six (nine seats), Forty-One (four seats), Forty-Six (twenty-seven seats — the late run’s fullest population), Forty-Seven (eight seats), and Fifty-One (fourteen standings, one in the full form “IEEE 754-2019”). At Twenty-Nine the IEEE name stands absent entire. None of these is a table position: the entry’s table positions remain two — Thirteen’s and Seventeen’s — and its grounded test values remain Thirteen’s rows.

The entry’s definition governs: the machine-precision states of the IEEE 754 binary formats as the books tabulate them — each format’s machine epsilon and exponent range. The reference reeled out of the entry is the table, citable, with the ε relation as its computed member; Book Seventeen’s table adds the quantization relation as a second member at its own place. One name, two standings, never confused: IEEE 754 is the floating-point standard of this page; IEEE Std 80, the grounding standard, is a different identity under the same institute’s name — it stands at its own calculator (the Grounding & Tolerable Voltage page), and in Fifteen and Sixteen the 754 identity is present exactly where the grounding apparatus is certified absent. Machine precision is also the floor beneath every computed figure on this site: the numerical substrate measuring the instruments themselves. Within SolveForce’s collective work the entry corresponds to the pillar SolveForce AI — the numerical precision of model computation — and, as a measured correspondence, to the arithmetic of digital communications (DSP, forward error correction, software radio), which runs in these formats; the book’s own placement is the computing substrate generally. No finder is served directly.

The printed states stand as printed, and the computed values stand beside them — nothing on this page is silently corrected. The digit states of record: Thirteen’s range minima print as the computed endpoints rounded at the stated precision (2−126 = 1.1755 × 10−38, printed 1.18 × 10−38; 2−1022 = 2.2251 × 10−308, printed 2.23 × 10−308), while its FP64 and FP128 range maxima print in the truncation state at the final digit (the computed endpoints 1.7977 × 10308 and 1.1897 × 104932, printed 1.79 × 10308 and 1.18 × 104932); Seventeen’s 10-bit Δq stands in the same truncation family at its row, as its panel states. The corpus records this digit-state family at its other seats (the Landauer printings; the semantic-compression quotients) — a state of the printed figure, carried beside the exact value, never overwritten by it.

For the figures: select a field, then tap the keys — or type as usual.