Oh, I didn’t know that…
Thank you for let me know 
I’m super happy that somebody does benchmarks. Good work phoronix. But I feel these test are mostly useless.
Who runs a 96 core CPU as a daily driver? There’s a huge bunch of scientific or server applications which are not super relevant for most people here.
How about a two year old mid-range desktop/laptop playing a youtube video with some simulated web surfing. Any difference in temperature or battery life? That would be interesting.
And then there are the details. Let’s take SVT-AV1: It has a heavy optimized v3 path which is always triggered, you don’t need to compile with v3. The v4 path on the other hand requires a build flag that isn’t set on the Arch package. Essentially every difference in the SVT-AV1 benchmark is benchmark hallucination.
2 Likes
hello , i have a trouble on update this weekend :
i get error on x265 and rubbeband , and theses points concerns Firefox , ffpmeg
waiting for return , update is possible
@BS86 can we get new build version on x265 / rubberband / firefox / ffmpepg / onevpl ?
Too bad you’ve deleted your comment. Somehow, as I haven’t played around with the ALHP repos yet, I was curious and installed the x86_64-v3 repos. I only have an Zen3 CPU in my desktop and there has been actually no real need for me to try these optimized builds.
So far, despite some issues with the gpg key for the repo being expired as of today, the switch was uneventful.
But I have to report that my GPU doesn’t behave that great with the optimized drivers, as I previously had an idle power consumption which was pretty low, those numbers went up and I’ve seen some dropped frames while playing back some youtube videos. Which hasn’t been an issue at all before the switch.
Maybe I’ll exclude the x86_64-v3 for the GPU driver stack as I was pretty happy with it before the switch. And the system at almost being idle without many workloads, just a browser being open and playing back some video is actually my major use-case. Thus, if the power consumption seems to be almost doubled in comparison to the regular repos… than that’s the route I’ll chase down.
1 Like
I was asking if anyone since the start of the thread had recent experience with the ALHP repo. If so how the experience was and if there was a noticeable difference for the user, I did some searching around later and seems no one is really able to come up with prove that you gain a noticeable difference with those repos enabled. People say the same about CachyOS giving better performance because of those optimized binaries, but I’ve decided just to stick with Arch default repos for now.
i check for alhp-keyring , they have to update because one key is missing
I was able to fix it by signing the key locally.
sudo pacman-key --lsign-key cie@harting.dev
I came across an interesting reply I thought I would share here since it seems the person that wrote the reply does seem to know what they are talking about. Here is the link and I’ll just quote the reply here.
v3 packages will not help you.
Every mathematical library and every machine learning application you’ve ever known already use the most advanced instruction sets available. They do so even when compiled for less capable architectures by detecting support at runtime, and they have done so since the beginning.
The only performance to be gained from v3 packages is the optimizations the compiler is able to find through autovectorization in the other parts of the compiled code. You’re not leaving “20 years of processor development on the table.”
It is expected that there is on average something to be gained from v3 packages, but nothing near what you are imagining, and certainly not for your workload.
So that makes me think if you won’t really notice a difference with v3(or v4) compiled packages with machine learning applications that would make me think you won’t really notice a difference either with normal desktop applications.
From different topics I have seen about this it just gives me the impression that it’s not worth using v3 or v3 packages from a third-party repo over using v2 packages. And same for as running a distribution like CachyOS even though people claim there is a performance improvement, but seems to me the performance improvement must come from else where?
There’s a huge performance difference with mathematical and machine learning applications. But because of that they often have runtime detection. So the answer is correct in the context of the question, which was about those topics.
But there are also applications that have compile time optimizations or that can be done by the compiler. There’s a reason everybody is coming out with v2/v3/v4 repos or make it the new baseline. It isn’t magic, but you leave a few percent on the table.
Are you going to notice it, probably not. Depends.
1 Like
That’s because with those applications it seems to be built into runtime, I might have worded wrong as in how I wrote my reply about it.
That’s the reason why I am not even going to try v3 or v4 packages or CachyOS, I’m getting the idea it’ll just be a disappointment in the performance gain I would be expecting over the reality of what gain you get.
As I’ve switched to the ALHP x86-64-v3 repos a week ago… I can confirm that I don’t see much differences in normal, daily desktop usage. But I’ve also haven’t pushed the system to its limits yet.
The switch was pretty unproblematic. And can be done by simply adding the ALHP mirrorlist and the keyring via yay / pacman. And adding the repos to your pacman.conf. With a subsequent system upgrade.
In short, my major objections are essentially two up so far:
-
As an advantage, local compile times when building packages from sources seems to be faster now.
-
Due to whatever reason the idle / light use power consumption of my GPU went up. I’m using a Radeon RX6600 which is configured to only run ast it’s lowest clock speeds usually. And from ~5 watt idle, it went up to ~15 watt. I’ll have to investigate a little why that’s the case.
Other than that, as the ALHP repos deploys a bit later than the regular arch repositories, they say their builds can take up to 24h or even more, depending on the complexity of the individual packages.
As a brief summary. Other than the GPU issue (which is a detail more or less), I haven’t experienced any issues due to the switch.
2 Likes