benchmarks

[Hugos68]

Sorry if this issue is inappropiate but I'm considering using tsdown instead of tsup/unbuild. But I wonder what improvements in speed this would result in. Are there any benchmarks of tsdown compared to tsup/unbuild/tsc, etc?

[sxzz]

Sure we'll prepare that in the upcoming docs #93

[Gugustinette]

@sxzz I started a benchmark setup here : https://github.com/Gugustinette/bundler-benchmark

For now I only made an exemple bundling a thousand functions, do you have ideas of what would be relevant or not ? I'm not much into bundlers performance.

Meanwhile I'll improve the benchmark architecture and automate a few graph etc...

[Gugustinette]

btw this are the numbers on the thousand function bundling, pretty cool !

[Gugustinette]

We can generate such graphs 👍

[Hugos68]

Damn, twice as fast as tsup! Nice!

[sxzz]

Good job! We also need dts comparing

[Gugustinette]

@sxzz I added a dts benchmark.

However, the generated index.d.mts file is empty using tsdown, while it's not with tsup and unbuild 🤔
That's all I get :

//# sourceMappingURL=index.d.mts.map

[Gugustinette]

I updated tsdown to latest version and re-ran the benchmark, this is what we have now :

What should we do @sxzz ? Value some numbers here on the documentation (similar to Rolldown docs) ?

We need to close this issue in some way lol

[sxzz]

I'm still working on improving performance, so please wait a bit longer (it may take until the dts plugin is ported to Rust).

[Gugustinette]

Sounds good to me ! 👍

[altano]

When compiling each package separately in a large monorepo, where each package is fairly small, you'll see much smaller gains (over tsc). I've been experimenting with tsdown and here's what I see for a ~200 line, single ts file, using isolated decl, generating everything (dts+.d.map+js.map):

➜  remark-mdx-toc git:(05-04-tsdown) ✗ rm -rf dist && hyperfine --warmup 1 node_modules/.bin/tsc
Benchmark 1: node_modules/.bin/tsc
  Time (mean ± σ):     476.9 ms ±  21.9 ms    [User: 1011.7 ms, System: 59.3 ms]
  Range (min … max):   457.2 ms … 513.8 ms    10 runs
 
➜  remark-mdx-toc git:(05-04-tsdown) ✗ rm -rf dist && hyperfine --warmup 1 node_modules/.bin/tsdown
Benchmark 1: node_modules/.bin/tsdown
  Time (mean ± σ):     328.9 ms ±  49.7 ms    [User: 421.4 ms, System: 61.2 ms]
  Range (min … max):   302.7 ms … 467.8 ms    10 runs

The gains are further reduced (as a relative %) when going through tools like pnpm run or turbo

I assume this is mostly node startup overhead at this point, and there's nothing tsdown can do, but I just wanted to point it out in case I'm hitting something unexpected.

I'm obviously comparing apples and oranges here though, since tsc isn't a bundler. But most of my packages are node based and don't need bundling, so the results are relevant to me.