Benchmarks

Benchmarks

The numbers below were measured on an AMD Ryzen 9 3950X (16C/32T, 3.49 GHz) running Windows 11 24H2, with .NET 10.0.6, RyuJIT AVX2, a Release build and BenchmarkDotNet 0.14 / 0.15. The full per-suite tables and their history are in BENCHMARKS.md.

Headline (current)

Scenario	v1.9.0 era	now	Why it moved
SVG render a 1 000-point line	66 µs	33.4 µs	render-path work v1.10–v1.13
SVG render with LTTB downsample (100 000 pts)	1.78 ms	0.81 ms	same
JSON round-trip (Figure ⇄ JSON)	40 µs	30.8 µs	per-series FromSeriesDto dispatch
PNG export via SkiaSharp (simple)	27 ms	12.9 ms	typeface cache + SKFontStyle leak fix
PDF export (simple / complex)	—	26.2 / 29.3 ms	first published PDF numbers
SIMD Vec.Sum / Vec.Mean on 100 000 pts	—	19 µs, 0 B
TransformBatch (AVX) on 100 000 pts	—	208 µs

Streaming

Benchmark	Result
RingBuffer.Append	40M ops/sec (single-writer, 100K capacity)
RingBuffer.ToArray (10K snapshot)	190K snapshots/sec
StreamingLine: 100K appends + snapshot	13 ms total

Scenario	Buffer	Append rate	Render FPS	Memory
Dashboard	1K	1/sec	1	16 KB
Telemetry	10K	100/sec	30	640 KB
Oscilloscope	100K	10K/sec	60	800 KB
Trading (OHLC)	5K	10/sec	10	160 KB

The ring buffer is never the bottleneck; rendering is. At 30 fps the snapshot takes under a microsecond, so each frame still has about 32 ms left.

Geographic projections (100K forward projections)

Projection	ops/sec		Projection	ops/sec
Sinusoidal	83M		Stereographic	24M
NaturalEarth	77M		EqualEarth	19M
Robinson	33M		Orthographic	17M
AlbersEqualArea	29M		PlateCarree	16M
Mercator	25M		LambertConformal	16M
			TransverseMercator	14M
			AzimuthalEquidistant	13M
			Mollweide	6M

Every projection exceeds 5M/sec. Projecting all 177 countries takes under 1 ms, even with Mollweide. Natural Earth 110m coastlines load in 5 ms (134 features). Countries load in 48 ms (177 features) and are then cached.

MathText and themes

Benchmark	Result
Parse \sum_{i=0}^{n} \frac{1}{i!} = e	473K parses/sec
Parse \begin{pmatrix} a & b \\ c & d \end{pmatrix}	554K parses/sec
30 theme presets loaded	370 µs

SVG rendering

Chart	Time	Allocated
Simple line (100 pts)	94 µs	136 KB
Line + scatter + bar	109 µs	133 KB
3×3 subplot grid	754 µs	933 KB
Treemap (6 nodes)	60 µs	109 KB
Sankey (4 nodes, 4 links)	63 µs	118 KB
Polar line (50 pts)	33 µs	56 KB
3D surface (10×10)	69 µs	124 KB
3D surface + directional lighting	82 µs	148 KB
Line + legend (3 series)	140 µs	214 KB
Large line (10K pts)	3.1 ms	3.7 MB
Large line (100K pts, LTTB→2K)	1.3 ms	2.4 MB

Technical indicators (100K points ≈ a trading day at 1-second bars)

Every indicator completes in under 3.3 ms. Multiple indicators run in parallel on separate cores.

Indicator	Time		Indicator	Time
SMA(20)	195 µs		MACD(12,26,9)	1.50 ms
VWAP	238 µs		ParabolicSAR	1.21 ms
EquityCurve	226 µs		CCI(20)	2.16 ms
EMA(20)	491 µs		BollingerBands(20)	2.23 ms
OBV	645 µs		ADX(14)	2.43 ms
RSI(14)	892 µs		WilliamsR(14)	2.97 ms
			Stochastic(14,3)	3.31 ms

Coordinate transform

The transform from data space to pixel space is the hot path. It is a single-pass AVX SIMD interleave: Vector256.Multiply and Add (FMA when available), then UnpackLow/High, then Permute2x128, then a store via MemoryMarshal.Cast. Off x86 it falls back to scalar code. It takes 764 ns for 1K points, 53 µs for 10K and 208 µs for 100K. Every line, scatter, area and bubble renderer uses this transform, so indicator output benefits from it automatically.

Why server-side SVG

Benefit	Detail
Zero client cost	the browser swaps innerHTML — no canvas redraw, no layout recalculation
Inline SVG	part of the DOM: CSS-styleable, screen-reader accessible, prints as vector
Consistent	every client sees the same chart, no browser rendering differences
Bandwidth	a typical chart is 5–15 KB; SignalR pushes only what changed
Scales	parallel subplot rendering uses every core

Run them yourself

cd Benchmarks/MatPlotLibNet.Benchmarks
dotnet run -c Release -- --filter "*SvgRendering*"
dotnet run -c Release -- --filter "*Indicator*"
dotnet run -c Release -- --filter "*"      # all suites

Run one suite at a time. Concurrent runs contend for the CPU and inflate the timings.