What is ASCII vs UTF8 vs UTF32 vs UTF16. Does it really matter?

ASCII (American Standard Code for Information Interchange) - 1 Byte

Originating in the 1960s, ASCII uses 7 bits to represent each character. This might prompt you to ask, "Why only 7?" Not every system had Byte Addressable Memory back in the day. And, a 7-bit length was adequate to signify the English alphabet (covering both lowercase and uppercase), numerals, punctuation marks, and select special control characters.

UTF (Unicode Transformation Format)

As ASCII had only English alphabets the need arose to depict other global characters.UTF tried to resolve that. It offers a method to encode characters as "code points." Typically, these code points are penned as U+XXXX, where 'XXXX' is a series of hexadecimal numbers.

UTF-8 (8-bit) - Ranges from 1 Byte to 4 Bytes

It is variable length encoding. UTF-8 can utilize sequences ranging from 1 to 4 bytes to signify each character. It is backward compatible with ASCII.

  • 1 byte: Traditional ASCII
  • 2 bytes: Encompasses Arabic, Hebrew, and the majority of European scripts
  • 3 bytes: Refers to the Basic Multilingual Plane* (BMP).
  • 4 bytes: Accounts for all Unicode characters.

utf 8

Let’s take an example.

The letter "A" (U+0041):

In UTF-8, the letter "A" is an ASCII character, which means it is encoded using a single byte. The binary representation of "A" in UTF-8 is as follows:

Binary: 01000001

Hexadecimal: 41

The Euro sign "€" (U+20AC):

The Euro sign "€" is a non-ASCII character with a higher Unicode code point. It is encoded using three bytes in UTF-8. The binary representation of the Euro sign in UTF-8 is as follows:

Binary: 11100010 10000010 10101100

Hexadecimal: E2 82 AC

In memory, the Euro sign "€" is represented using three consecutive bytes with the hexadecimal values E2, 82, and AC.

UTF-32 (32-bit) - A fixed 4 Bytes

This encoding depicts every Unicode character with a direct 4-byte (32-bit) representation. 4 bytes: Every character, ranging from U+0000 to U+10FFFF, gets a fixed representation. This makes applying string operation very easy.

UTF-16 (16-bit) - Spans 2 Bytes to 4 Bytes

Another variable-length encoding, UTF-16 predominantly employs sequences of either 1 or 2 units of 16 bits each.

2 bytes: This engulfs the BMP, representing a gamut of characters from U+0000 to U+FFFF.

4 bytes: Characters outside the BMP, covering characters from U+010000 to U+10FFFF.

Asian Char

Now you might ask why we can't represent ‘あ’ in 2byte in UTF-8. See in UTF-8 Encoding Structure the 2-byte sequences are of the form 110xxxxx 10xxxxxx. We are wasting 5 bits out of 16 bits to make it compatible for variable-length encoding. We don’t have that issue with UTF16. So the character dictionary increases.


  1. UTF-8 is Size Optimized
  2. UTF-32 is Perfomance Optimized
  3. UTF-16 Tries to strike a balance between the two.

Choosing the right text encoding can significantly influence storage efficiency, interoperability, and compatibility. For instance, software designed for ASCII might malfunction when faced with UTF-8 data it wasn't expecting.

The choice between UTF-8 and UTF-16 can also impact the performance of training large language models, especially when dealing with Asian characters. In UTF-8, many Asian characters take up 3 bytes, while in UTF-16, they usually occupy just 2 bytes. This difference might seem trivial, but it adds up when training on vast datasets. The more bytes required for each character in UTF-8, the longer the sequences, which means more computation during training. Longer sequences also demand more memory.

So, for languages rich in Asian characters, UTF-16 could offer a more efficient training route, thanks to its relatively compact representation compared to UTF-8.