Kibibyte to Exbibyte

KB (kilobyte, SI) = 1,000 bytes. KiB (kibibyte, IEC binary) = 1,024 bytes. The difference is 24 bytes (2.4%) — small individually but the source of the well-known discrepancy between storage manufacturer labels and OS-reported sizes. Storage manufacturers use KB = 1,000 bytes; operating systems traditionally used KB = 1,024 bytes (now correctly called KiB).

Linux memory management, filesystem block sizes, and page sizes are all powers of 2 (typically 4,096 bytes = 4 KiB). Using kibibytes aligns with the physical hardware structure. The GNU coreutils (df, du, ls -h) display sizes in KiB, MiB, GiB by default for consistency with how the kernel allocates memory and disk blocks — decimal kilobytes would produce fractional values for normal aligned allocations.

Most languages expose both conventions depending on the API. Java's Runtime.totalMemory() returns bytes aligned to KiB (binary), but Files.size() returns raw byte counts that file managers may display as decimal KB. Python's os.path.getsize() returns bytes — the developer chooses how to format. Go's humanize library defaults to IEC (KiB) while many JavaScript libraries default to SI (KB). This inconsistency means the same file can appear as different sizes across tools written in different languages.

A memory page is the smallest unit of memory the OS allocates from physical RAM. Most modern CPUs use 4 KiB (4,096 byte) pages; some support 2 MiB or 1 GiB "huge pages" for performance. Every memory allocation is rounded up to the nearest page boundary. This binary alignment is why computer memory sizes are always powers of 2 (4 GB, 8 GB, 16 GB RAM) rather than round decimal numbers (5 GB, 10 GB).

The 3.5-inch floppy's capacity was 1,474,560 bytes — which is neither 1.44 MB (1,440,000 bytes) nor 1.44 MiB (1,509,949 bytes). The label came from a hybrid calculation: 80 tracks × 2 sides × 18 sectors × 512 bytes = 1,474,560 bytes, then divided by 1,000 to get 1,474.56 KB, then divided by 1,024 to get "1.44 MB." This mix of decimal and binary division in the same label is one of the most famous unit blunders in computing history.

EB (exabyte) = 10¹⁸ bytes (SI decimal). EiB (exbibyte) = 2⁶⁰ bytes = 1,152,921,504,606,846,976 bytes (IEC binary). EiB is 15.29% larger. This is the largest practically significant SI vs IEC discrepancy: per exbibyte, the binary value exceeds the decimal value by approximately 152,921,504,606,846,976 bytes — about 152.9 petabytes.

One exbibyte (EiB) ≈ 1.153 × 10¹⁸ bytes = 1,073,741,824 GiB = 1,048,576 TiB. In practical terms: enough to store approximately 230 billion JPEG photos at 5 MB each, or 288,230,376 copies of a 4 GB HD movie, or the entire text content of the English internet many thousands of times over.

In theory, yes — and with astonishing density. DNA can encode about 215 PiB per gram of material, meaning a single EiB could fit in roughly 4.7 grams of synthetic DNA. Researchers at Microsoft and the University of Washington have demonstrated writing and reading megabytes of data in DNA strands. The challenges are speed and cost: current DNA synthesis writes about 400 bytes per second and costs around $3,500 per megabyte. At that rate, writing 1 EiB would take billions of years and cost more than global GDP. However, enzymatic synthesis breakthroughs could reduce costs by 6–8 orders of magnitude within decades.

Storing 1 EiB on modern HDDs would require roughly 57,000 drives of 20 TB each, consuming about 400–500 kW of power just for the drives — plus 200–300 kW for cooling, networking, and overhead. That totals roughly 6 GWh per year, equivalent to powering about 550 US homes. At typical US grid carbon intensity, this produces around 2,500 tonnes of CO₂ annually. Hyperscale operators reduce this via renewable energy and immersion cooling, but the fundamental physics of spinning magnetic platters or maintaining NAND charge states sets a floor on energy consumption that no software optimisation can eliminate.

After exbibyte (EiB, 2⁶⁰ bytes) come: zebibyte (ZiB, 2⁷⁰ bytes) and yobibyte (YiB, 2⁸⁰ bytes), as defined in IEC 80000-13. These are recognized standard units but have no current practical applications. The entire global internet's estimated stored data (hundreds of EB) is still in the low hundreds of EiB range — well short of one ZiB.