Is it expected that `mozjpeg: true` uses significantly more cpu?

[styfle]

Question about an existing feature

Is it expected that setting mozjpeg: true and processing jpeg as both input and output uses significantly more cpu than the default?

What are you trying to achieve?

Trying to find a good default for Next.js users. I noticed really high CPU for JPEG input and output.

In a previous PR, @lovell set mozjpeg to always be true here:

• fix(next/image): prefer sharp defaults, use mozjpeg for JPEG vercel/next.js#65846

When you searched for similar issues, what did you find that might be related?

https://github.com/lovell/sharp/issues?q=is%3Aissue%20state%3Aclosed%20mozjpeg

Please provide a minimal, standalone code sample, without other dependencies, that demonstrates this question

TODO

Please provide sample image(s) that help explain this question

2,149 JPEGs tested with MozJPEG on and off at quality 75 with concurrency of 8 on MacBook M3 Pro.

CPU averages use each image’s median of three runs on an M3 Pro, including decode/rotate/resize/encode. Ratios average each image’s input KB ÷ output KB; higher means greater size reduction. Smaller images weren’t enlarged.

Requested width	Average CPU (ms)	CPU saved	Average output KB	Average input/output ratio
3840	31.67 → 8.80	72.2%	39.57 → 51.81	4.13 → 3.81
1920	20.96 → 7.72	63.2%	25.90 → 33.52	4.51 → 4.09
1080	14.48 → 5.76	60.2%	17.71 → 22.71	5.62 → 4.93
640	11.01 → 5.02	54.4%	12.76 → 16.31	8.31 → 7.09

Disabling MozJPEG saved 54–72% CPU, with average outputs 28–31% larger.

[lovell]

Yes, I enabled mozjpeg in vercel/next.js#65846 to ensure feature-parity with squoosh, which always used mozjpeg.

If CPU cost is now of more concern than bandwidth then either switch off mozjpeg or perhaps look at selectively disabling optimiseScans (whilst keeping mozjpeg enabled).

[styfle]

Thanks! I tried with both mozjpeg: true and optimizeScans: false and there was some improvement so that might be a good compromise since there's barely a size increase.

Width	Average CPU (ms)	CPU saved	Average output KB	Size increase
3840	32.03 → 23.76	25.8%	39.57 → 41.29	4.3%
1920	21.24 → 16.46	22.5%	25.90 → 26.82	3.6%
1080	14.66 → 11.60	20.8%	17.71 → 18.35	3.6%
640	11.18 → 9.13	18.4%	12.76 → 13.22	3.6%

It looks like mozjpeg sets about 5 different settings so I might have to try to benchmark all the combinations.

The thing to note about squoosh was that we always used sharp for production builds (next build) and squoosh was only used for development builds (next dev). Now of course, sharp is used for both.

I'm tempted to just turn of mozjpeg completely since it seems like no one noticed the difference between on vs off.

How did you decide the default settings for sharp?

[lovell]

How did you decide the default settings for sharp?

The default is all mozjeg features off, which is the same as using libjpeg-turbo (upon which mozjpeg is based), as that's what sharp provided for many years.

Switching mozjpeg on is syntactic sugar to set all the mozjpeg-related options to their mozjpeg defaults:

sharp/lib/output.mjs

Lines 637 to 643 in 9dcb3dd

	if (options.mozjpeg) {
	this.options.jpegTrellisQuantisation = true;
	this.options.jpegOvershootDeringing = true;
	this.options.jpegOptimiseScans = true;
	this.options.jpegProgressive = true;
	this.options.jpegQuantisationTable = 3;
	}