AMD Ryzen Threadripper 3990X Linux Performance After Three Years

[Linuxxx]

Strange how video encoding seems to be degraded over all codecs but basically everywhere else improvements.

That's just the magic of the schedutil governor at full-force!

When it can't even beat ondemand, why even bother?

Oh that's right, upstream already stopped caring years ago!

Too bad that the Steam Deck is still defaulting to it, among many other Linux distributions...

[qarium]

Perhaps he's just jealous of Michael's hot girlfriend/wife?
On-topic: it is amazing how much more performance can be squeezed out of an older platform. It means that launch benchmarks are not the end of all. I'm curious how Zen4 will improve on Linux platform the coming years...

is this true for the cpu ?... the cpu side by fixing all the security holes rarely goes faster.

all the big jumps in performance is the GPU side and thats only because 'AMD sells its hardware with unfinished driver compared to this Nvidia only sells hardware with matured drivers and because of this most of the time do not see any speedup.,

[onlyLinuxLuvUBack]

there is also thermal paste it looks like the tissues is to remove unwanted thermal paste...

he doesn't blow his nose with it.

thats bs , everyone knows to use coffee filters to wipe paste, and larabel is using tissue for the cute cat pics

[Ray Ingles]

Minor typo: "The Thelio chassis still looks as good now as it did back in 2023..."

[Alexmitter]

Could Mister Larabel at least put away his box of you-know-what tissues conveniently situated next to the computer before taking photos of it? Some of us don't want to imagine how they're being used while reading the article. Gross!

Some people have noses.

[rabcor]

This is incredible, such improvement, much wow

Could Mister Larabel at least put away his box of you-know-what tissues conveniently situated next to the computer before taking photos of it? Some of us don't want to imagine how they're being used while reading the article. Gross!

No can do, he gets paid for the tissue ad

Kimtech! The tissue of choice for linux nerds!

[sophisticles]

The video encoding benchmark results should indicate a major problem with the testing methodology.

Video encoders, x264, x265 and the svt family of encoders have had their settings refactored a number of times over the years.

For instance, if you used the default settings, x264 used to default to the medium preset, then after a refactoring of settings, the old medium preset became the new fast preset and different internal configurations that used to be the default for each preset were rearranged.

Another example is AQ, x264 used to only enable AQ after a certain preset, now it's always enabled, over the years they changed the default from AQ 1 to AQ 2, MB-Tree iirc is now always enabled.

Long story short, if this test is comparing benchmark results obtained in 2020 using the latest builds of these encoders available in 2020 against results obtained in 2023 using the latest builds available today, then the results are invalid, these encoders have had significant changes to their code and default settings and while the encoding may be slower I'm guessing the quality of the encodes has gone up.

This is the problem with relying on so-called automated testing to run benchmarks that you really don't understand, you get nonsense results that mislead your readers.

[Michael]

The video encoding benchmark results should indicate a major problem with the testing methodology.

Video encoders, x264, x265 and the svt family of encoders have had their settings refactored a number of times over the years.

For instance, if you used the default settings, x264 used to default to the medium preset, then after a refactoring of settings, the old medium preset became the new fast preset and different internal configurations that used to be the default for each preset were rearranged.

Another example is AQ, x264 used to only enable AQ after a certain preset, now it's always enabled, over the years they changed the default from AQ 1 to AQ 2, MB-Tree iirc is now always enabled.

Long story short, if this test is comparing benchmark results obtained in 2020 using the latest builds of these encoders available in 2020 against results obtained in 2023 using the latest builds available today, then the results are invalid, these encoders have had significant changes to their code and default settings and while the encoding may be slower I'm guessing the quality of the encodes has gone up.

This is the problem with relying on so-called automated testing to run benchmarks that you really don't understand, you get nonsense results that mislead your readers.

The tests are version-locked to the given test result. The same version of x264 / x265 / svt-av1 were used between the 2020 and 2023 comparisons. The video encode performance hit is likely due to the CPU frequency scaling driver/governor changes.

[Anux]

That's just the magic of the schedutil governor at full-force!

I'm just thinking about how much more improvements we could see, if the test weren't done with a defect governor.