.NET CSV benchmark

Benchmark source
Test last run 05/01/2026

Libraries compared:

Sylvan.Data.Csv v1.4.3

Sep v0.12.1

FourLambda.Csv v1.1.0

NReco.Csv v1.0.3

CsvHelper v33.1.0

reddit_subset.csv

5.2 MB (UTF-8), 10.5 MB (UTF-16)[22 columns, 16,781 rows]

'RC_2008-01.ndjson' from the Arctic Shift reddit dump converted to CSV format. Meant to represent real-world data processing, and contains a very diverse range of long, complex text.

mixed_unicode.csv

5.5 MB (UTF-8), 8.0 MB (UTF-16)[4 columns, 52,524 rows]

'item_flavor_text.csv' from the Veekun/pokedex GitHub repo. Contains medium-length text in many different languages & scripts, with many newlines.

65K_Records_Data.csv

7.9 MB (UTF-8), 15.7 MB (UTF-16)[14 columns, 65,535 rows]

Example CSV dataset used by MarkPflug/Benchmarks to benchmark .NET CSV libraries. Contains sample financial record data.

synthetic_thin_numeric.csv

916.9 KB (UTF-8), 1.8 MB (UTF-16)[4 columns, 50,000 rows]

An artificially generated dataset consisting of 4 columns containing integers ranging from 0 - 2000.

synthetic_short_numeric.csv

969 B (UTF-8), 1.9 KB (UTF-16)[4 columns, 50 rows]

An artificially generated dataset consisting of 4 columns containing integers ranging from 0 - 2000. Short to show overall library overhead.

synthetic_quoted_numeric.csv

1.3 MB (UTF-8), 2.6 MB (UTF-16)[4 columns, 50,000 rows]

An artificially generated dataset consisting of 4 columns containing integers ranging from 0 - 2000, with each column wrapped in quotation marks for escaping.

synthetic_complex.csv

70.3 MB (UTF-8), 134.3 MB (UTF-16)[4 columns, 50,000 rows]

An artificially generated dataset consisting of 1 column with a number, and 3 containing very large escaped text designed to be extremely challenging for a parser to scan through.

Benchmark environment:

BenchmarkDotNet v0.15.8, Windows 10 (10.0.19044.6575/21H2/November2021Update)
AMD Ryzen 7 9700X 3.80GHz, 1 CPU, 16 logical and 8 physical cores
.NET SDK 10.0.101
  [Host] : .NET 10.0.1 (10.0.1, 10.0.125.57005), X64 RyuJIT x86-64-v4

Tests come in two variants:

UTF-8: Loaded binary CSV as UTF-8 data read from a Stream. Sylvan.Data.Csv does not directly support this, and is emulated by passing it through a StreamReader.
UTF-16: Loaded CSV as UTF-16 data read from a TextReader.
If you parse a CSV with the source as a string / Span<char>, you will be using this version.
Note that while libraries using this seem to have faster raw MB/s speed, it's not necessarily processing more data faster; on average UTF-16 versions of UTF-8 data is double the physical byte size. Use Mean as an unbiased reference if you want to measure actual parse time

Test suite #1 - Skip to end

Measures how fast a parser can read from start to finish, without extracting any values and only interpreting structure. Mimics workloads like:

Counting the amount of lines and fields in a CSV
Validating column count

UTF8

Library		Speed		Mean	Error	Allocations
#1	FourLambda.Csv	4,213.46 MB/s	±3.84 MB/s	1,245.140 µs	1.135 µs	0.64 KB
#2	Sep	3,934.85 MB/s	±1.77 MB/s	1,333.304 µs	0.600 µs	6.58 KB
#3	Sylvan.Data.Csv	1,088.63 MB/s	±1.12 MB/s	4,819.209 µs	4.976 µs	77.04 KB
#4	NReco	1,047.41 MB/s	±2.37 MB/s	5,008.895 µs	11.352 µs	84.91 KB
#5	CsvHelper	565.40 MB/s	±0.62 MB/s	9,279.084 µs	10.172 µs	72.28 KB

UTF16

Library		Speed		Mean	Error	Allocations
#1	Sep	8,160.25 MB/s	±20.52 MB/s	1,285.392 µs	3.233 µs	3.28 KB
#2	FourLambda.Csv	6,954.56 MB/s	±22.80 MB/s	1,508.236 µs	4.946 µs	0.62 KB
#3	Sylvan.Data.Csv	2,167.21 MB/s	±4.30 MB/s	4,839.910 µs	9.593 µs	74.08 KB
#4	NReco	2,137.50 MB/s	±5.70 MB/s	4,907.190 µs	13.092 µs	81.61 KB
#5	CsvHelper	1,144.64 MB/s	±1.58 MB/s	9,163.673 µs	12.650 µs	68.98 KB

Test suite #2 - Column sum

Measures how fast a single column can be parsed as an int32 and compute a sum. Mimics workloads like:

Calculating a sum / min / max across transactions
Determining the maximum ID of a dataset of records

UTF8

Library		Speed		Mean	Error	Allocations
#1	FourLambda.Csv	3,934.93 MB/s	±9.60 MB/s	1,333.275 µs	3.253 µs	0.65 KB
#2	Sep	3,626.02 MB/s	±3.55 MB/s	1,446.862 µs	1.416 µs	6.57 KB
#3	Sylvan.Data.Csv	1,072.36 MB/s	±1.41 MB/s	4,892.356 µs	6.437 µs	77.04 KB
#4	NReco	1,003.88 MB/s	±2.74 MB/s	5,226.069 µs	14.286 µs	500.26 KB
#5	CsvHelper	519.84 MB/s	±1.07 MB/s	10,092.176 µs	20.681 µs	1,274.87 KB

UTF16

Library		Speed		Mean	Error	Allocations
#1	Sep	7,551.89 MB/s	±8.87 MB/s	1,388.939 µs	1.631 µs	3.28 KB
#2	FourLambda.Csv	6,313.79 MB/s	±12.83 MB/s	1,661.305 µs	3.376 µs	0.62 KB
#3	Sylvan.Data.Csv	2,149.64 MB/s	±2.03 MB/s	4,879.480 µs	4.611 µs	74.08 KB
#4	NReco	2,033.50 MB/s	±2.36 MB/s	5,158.170 µs	5.991 µs	496.96 KB
#5	CsvHelper	1,047.42 MB/s	±1.51 MB/s	10,014.246 µs	14.406 µs	1,271.57 KB

Test suite #3 - Read all columns

Measures how fast a parser can get the text content of every column. Has 3 variants:

"string": Reads the column value as a string object
"span": Reads the column value as a Span<char> object. Not available in the UTF-8 version of FourLambda.Csv
"span-raw": Reads the column value as a Span<char> / Span<byte> object, without performing any unescaping. Not available in Sylvan.Data.Csv

string / UTF8

Library		Speed		Mean	Error	Allocations
#1	Sep	1,031.48 MB/s	±5.06 MB/s	5,086.238 µs	24.972 µs	16.15 MB
#2	FourLambda.Csv	671.82 MB/s	±2.42 MB/s	7,809.193 µs	28.140 µs	17.93 MB
#3	Sylvan.Data.Csv	576.33 MB/s	±2.03 MB/s	9,102.983 µs	32.021 µs	18.11 MB
#4	NReco	555.64 MB/s	±2.63 MB/s	9,442.045 µs	44.671 µs	20.23 MB
#5	CsvHelper	357.35 MB/s	±0.94 MB/s	14,681.397 µs	38.818 µs	18.06 MB

string / UTF16

Library		Speed		Mean	Error	Allocations
#1	Sep	2,088.17 MB/s	±7.70 MB/s	5,023.126 µs	18.532 µs	16.14 MB
#2	FourLambda.Csv	2,007.17 MB/s	±8.93 MB/s	5,225.835 µs	23.241 µs	17.93 MB
#3	Sylvan.Data.Csv	1,161.66 MB/s	±3.02 MB/s	9,029.427 µs	23.448 µs	18.11 MB
#4	NReco	1,123.55 MB/s	±5.16 MB/s	9,335.678 µs	42.913 µs	20.23 MB
#5	CsvHelper	718.35 MB/s	±1.87 MB/s	14,601.614 µs	38.018 µs	18.06 MB

span / UTF8

Library		Speed		Mean	Error	Allocations
#1	Sep	2,311.40 MB/s	±0.42 MB/s	2,269.776 µs	0.411 µs	6.59 KB
#2	FourLambda.Csv	1,142.12 MB/s	±1.50 MB/s	4,593.532 µs	6.050 µs	0.68 KB
#3	Sylvan.Data.Csv	900.14 MB/s	±1.18 MB/s	5,828.399 µs	7.629 µs	178.52 KB

span / UTF16

Library		Speed		Mean	Error	Allocations
#1	FourLambda.Csv	5,240.63 MB/s	±3.60 MB/s	2,001.500 µs	1.375 µs	0.63 KB
#2	Sep	4,698.56 MB/s	±7.60 MB/s	2,232.410 µs	3.613 µs	3.29 KB
#3	Sylvan.Data.Csv	1,790.83 MB/s	±4.93 MB/s	5,857.138 µs	16.128 µs	175.56 KB

span-raw / UTF8

Library		Speed		Mean	Error	Allocations
#1	FourLambda.Csv	3,809.30 MB/s	±1.69 MB/s	1,377.249 µs	0.612 µs	0.65 KB
#2	Sep	3,395.33 MB/s	±2.79 MB/s	1,545.167 µs	1.269 µs	6.57 KB

span-raw / UTF16

Library		Speed		Mean	Error	Allocations
#1	Sep	7,059.28 MB/s	±20.21 MB/s	1,485.862 µs	4.253 µs	3.28 KB
#2	FourLambda.Csv	6,380.35 MB/s	±19.20 MB/s	1,643.972 µs	4.948 µs	0.62 KB

Test suite #4 - Composite workload

Measures how fast a parser can extract strongly-typed values from rows, e.g. extracting a decimal from a decimal formatted column, and a DateTime from a datetime formatted column.

Not available for this dataset