JSON Parser Benchmarks

Cross-language throughput comparison of 16 JSON parsers. Numbers in MB/s, sorted by parse speed.

ParserLanguageParse (MB/s)Stringify (MB/s)Notes
simdjsonC++
4,200
—SIMD AVX-512. Reference for what is possible.
simd-json (Rust)Rust
2,400
—SIMD port of simdjson. Requires mutable input buffer.
sonic-cppC++
2,100
1,900ByteDance, SIMD-accelerated.
msgpackrJavaScript
720
850Same author as cbor-x. Compatible with both MessagePack and JSON.
sonic-goGo
580
720JIT-compiled. Drop-in replacement for encoding/json.
orjsonPython
550
900Rust-based. Default choice for performance-sensitive Python.
simd-json (Python)Python
480
—simdjson bindings via Cython.
json-iteratorGo
420
520Drop-in for encoding/json, lookup-table based.
serde_jsonRust
380
420Universal default in Rust. Predictable, well-tested.
System.Text.JsonC#
380
440Built-in since .NET 5. Faster, smaller allocations.
ujsonPython
320
410C-based, popular before orjson.
V8 JSON.parseJavaScript
280
340Node 20 / Chrome. The default everyone uses.
JacksonJava
280
330De facto Java default.
Newtonsoft.JsonC#
220
260Legacy default. Beaten by System.Text.Json.
encoding/jsonGo
110
180Standard library. Reflection-based, hence slower.
json.dumps (stdlib)Python
90
130Pure Python. Baseline reference.

How to read this

The numbers are end-to-end throughput on a single core for medium-size documents (~600 KB) of mixed shape — the standard simdjson benchmark inputs twitter.json and citm_catalog.json. Real-world performance varies 10-20% depending on CPU, payload shape, and exact library version.

simdjson is the ceiling — what is possible when you use SIMD instructions and design the entire parser around them. Most production JSON parsing is at 200-400 MB/s, which is fine for everything except very high-throughput data pipelines.

Pick by language ergonomics first. If your codebase is Go, encoding/json at 110 MB/s is probably acceptable. If parsing becomes the bottleneck, swap to sonic or json-iterator — same API, different engine. Premature optimization here usually costs more than it saves.

trending_up

When raw speed matters

High-throughput pipelines (Kafka consumers, log shippers, telemetry ingestion). Profile first — most other JSON code lives elsewhere.

balance

When it does not

APIs where the network round-trip dominates. CLI tools that parse one document. Anything where parse time is < 5% of CPU.

FAQ

Where do these numbers come from?
Published benchmarks from each project's README plus the simdjson benchmarks suite (twitter.json, citm_catalog.json, ~600KB inputs). Measurements vary 10-20% by CPU, payload shape, and exact version — treat the table as relative ordering, not absolute truth.
Why is my real-world parsing slower?
Three reasons: cold parser (first call pays JIT compilation), payload shape (deeply nested objects are slower than flat arrays of records), and downstream work (allocation, schema validation, copying values into typed structs). End-to-end throughput is usually half what the parser claims.
Should I switch to simdjson?
Only if parsing is your bottleneck. For most APIs, the network and the business logic dominate. Profile first. simdjson requires the input as a mutable byte buffer it can write back into — not a drop-in replacement for the standard parser.
What about streaming?
These numbers are for full-document parsing. For streaming (NDJSON, partial JSON, multi-GB documents) the relevant parsers are stream-json (Node), ijson (Python), encoding/json.Decoder (Go) — they trade peak throughput for constant memory.