AMD Ryzen AI 300 Series Dominates Intel Core Ultra 7 Lunar Lake Performance For Linux Developers & Creators

[Nempk]

You're completely off-target: this processor is not intended to compete as a top performer, otherwise it would not be a 17W part.
This new architecture and its fabrication process are promising, the efficiency graphs show a great leap and the gap against AMD is greatly reduced. While I expect the gap to widen a bit with higher power chips (but it could even not be the case), this is indeed a valid architecture.

Furthermore: Intel will not go bankrupt for this. They still have something like 80% of the market, they sold crap until yesterday for years and did not go bankrupt, they won't go bankrupt today nor tomorrow.

The only problem for this chips from intel is the price. I can't imagine anyone who really use the computer for more than watching videos and using text editor choosing this. The price is the most important factor, unless they prefer blue sticker more than red(can be the case)

[69Y69]

The heterogeneous scheduling of Strix Point SoCs doesn't work well under Linux at the moment. I've been daily driving the very same laptop tested in the article for several days using latest 6.11 kernel with all known kernel patches applied, and it still priotizes Zen5c cores for a lot of the single-threaded scenarios.

This is sad to read. Heterogenous scheduling is a hard problem to solve, yet we are spammed by E-cores by both x86 vendors.

Anyways, have you tried these recent Detect max performance values for heterogeneous AMD designs patches?

[David Huang]

This is sad to read. Heterogenous scheduling is a hard problem to solve, yet we are spammed by E-cores by both x86 vendors.

Anyways, have you tried these recent Detect max performance values for heterogeneous AMD designs patches?

Yes, I immediately applied that when I saw the report from Phoronix. Unfortunately it doesn't fix the problem.

Actually the whole "preferred cores" scheduling of AMD Ryzen seems broken at least on multi-CCD chips. I own multiple Ryzen desktop and laptop systems and none of them prioritizes highest-perf preferred cores properly. I do occasionally get it to work on single-CCD systems, but that's not my primary use case.

In theory HX 370 is nothing special, it's just an extended version of preferred core scheduling. But the whole thing is not in a good shape at the moment for Linux on Ryzen.

[chithanh]

It's possible there is something quirky going on with the ASUS Zenbook S 14 and its firmware/BIOS under Linux comparative to Windows, but that isn't yet confirmed and nothing more than speculating what may be hindering Lunar Lake on Linux.

One way to tell would be to compare WSL vs. native Linux performance, once on the ASUS Zenbook S 14 and once on some other Laptop which is known to deliver expected performance like the Zenbook S 16 with the Ryzen 365/HX370.

If WSL performs better than native Linux on Lunar Lake relative to how WSL performs vs. native Linux on Strix Point, then it is probably some Linux-specific problem with platform support.

[Kabbone]

I recently bought a Lunar Lake laptop myself (MSI Prestige 13 AI+ EVO w/ an Intel Core Ultra 7 258V - boy that naming is awful) and these performance results don't surprise me that much - LNL has 4 P-Cores locked at 17W. But, I mainly got it because it was a great ultralight config at <1kg of weight with a 75Wh battery. My powertop and powerstat testing has whole laptop idling as low as 2.3W and under light use (text editing, browsing) in GNOME it seems to hang around 5-8W, which is not bad.

For me a bunch of stuff (GPU, WiFi) didn't work well w/ 6.11, so I had to go to 6.12rc1 mainline - I do think the Linux support overall is undercooked overall. Also, the suspend on my laptop is wonky. When it does work, it still burns about 0.9% battery/h (almost 30% battery/day), doesn't seem to ever get to PC10 and running S0ixSelftestTool, only ever gets to S0i2.1. That isn't a dealbreaker, assuming suspend-then-hibernate can work, but it also seems to have intermittent RCU timeouts on resume that causes the laptop to immediately go back to suspend and never wake up (not great). Just normal Linux laptop things I guess, but hopefully some kernel updates can iron that out.
...

Interesting, how did you test the idle state, like what was still running?
By coincidence I just tested my 11th gen laptop and got it down to 2,9W (brightness 10%) or so but basically nothing running, even no Wifi.

But it's also bad with S0idle, it doesn't really work, like gpu crashing when waking up if it went lowest power state.

[willmore]

Funny you start by being toxic and offensive, then jump to arguing with me while heavily distorting facts to fit your "I buy laptops for heavy MT tasks, so should other people, thus AMD rules, Intel sucks, and I couldn't care less about battery life".

You were treated with more respect than you deserve. You made nonsensical and uninformed statements.

[randomfoo]

Interesting, how did you test the idle state, like what was still running?
By coincidence I just tested my 11th gen laptop and got it down to 2,9W (brightness 10%) or so but basically nothing running, even no Wifi.

But it's also bad with S0idle, it doesn't really work, like gpu crashing when waking up if it went lowest power state.

I tested the idle in TTY (with GDM running in the bg) and running `powerstat`. It's Arch-based so there's very little running in the background atm (mostly systemd stuff) but I did install `tuned`, `thermald`, and `auto-cpufreq`. `powertop` reports that all tuning parameters are 'GOOD'. What's somewhat impressive w/ the idle power is that I didn't turn the brightness down (I believe it was 50%), but it's an OLED display, so with a black TTY should be favoring lower power consumption. Wifi was left on. Note, my old Framework laptop had a 1260P that could also idle quite low, as low as 2.6W in TTY w/ 0% backlight, but it was also at ~99.8% C10 when idling whereas the new LNL chip doesn't seem to get below C7 atm (from `turbostat` results, `powerstat` only reports ACPI states (C1_ACPI-C3_ACPI) but it `powerstat -d 0 -c -H 1 480 -R -D` shows the CPU averaging 1.61W. Here's the breakdown from the RAPL reports:

Uncore	Package	Core	DRAM	Platform
0.00	0.60	0.02	0.03	0.95

Pretty impressive if I'm reading them correctly.

For your S0idle, I'd recommend trying out the S0ixSelftestTool (on Github) if you haven't yet, it might tell you something useful.

Oh also, in the BIOS I found a "Performance" setting which seems to disable the PL limits from `turbostat`'s perspective. In practice, this seems to take the power from being locked at 17W to about 29.5W.

The fans get quite loud, but temps on the laptop hold steady at about 84C. Anyway, at this point, if I could figure out the intermittent resume issues I'd be a pretty happy camper.

[Kabbone]

I tested the idle in TTY (with GDM running in the bg) and running `powerstat`. It's Arch-based so there's very little running in the background atm (mostly systemd stuff) but I did install `tuned`, `thermald`, and `auto-cpufreq`. `powertop` reports that all tuning parameters are 'GOOD'. What's somewhat impressive w/ the idle power is that I didn't turn the brightness down (I believe it was 50%), but it's an OLED display, so with a black TTY should be favoring lower power consumption. Wifi was left on. Note, my old Framework laptop had a 1260P that could also idle quite low, as low as 2.6W in TTY w/ 0% backlight, but it was also at ~99.8% C10 when idling whereas the new LNL chip doesn't seem to get below C7 atm (from `turbostat` results, `powerstat` only reports ACPI states (C1_ACPI-C3_ACPI) but it `powerstat -d 0 -c -H 1 480 -R -D` shows the CPU averaging 1.61W. Here's the breakdown from the RAPL reports:

Uncore	Package	Core	DRAM	Platform
0.00	0.60	0.02	0.03	0.95

Pretty impressive if I'm reading them correctly.

For your S0idle, I'd recommend trying out the S0ixSelftestTool (on Github) if you haven't yet, it might tell you something useful.

Oh also, in the BIOS I found a "Performance" setting which seems to disable the PL limits from `turbostat`'s perspective. In practice, this seems to take the power from being locked at 17W to about 29.5W.

The fans get quite loud, but temps on the laptop hold steady at about 84C. Anyway, at this point, if I could figure out the intermittent resume issues I'd be a pretty happy camper.

I think these are quite some impressive numbers if it fits your needs.

Thanks for the hints, I actually used S0idle before, at least I thought so, but it seems I didn't really look at it in detail because I didn't see the output without ssh'ing in and attach the tmux session. Turns out, it doesn't reach any S0idle state because some required IPs couldn't reach it:

PCIe/USB3.1_Gen2PLL_OFF_STS
SPC_PG_STS
MPHY_Core_DL_REQ_STS_16

Also the plugged in USB security key seemed to have been a problem. Need to research a bit how can overcome all this or I just need to hibernate.