Benchmarking the different machines in my office with the wonderful kcbench: http://www.kroah.com/log/blog/2025/10/01/the-only-benchmark-that-matters-is.../

@gregkh AMD EPYC 9754 128-Core Processor [512 threads]
Cpufreq; Memory: Unknown; 773414 MiB
Compiler: gcc (Debian 15.2.0-4) 15.2.0
Linux compiled: 6.17.0 [/data/linux-6.17/]
Config; Environment: defconfig; CCACHE_DISABLE="1"
Build command: make vmlinux
Run 1 (-j 512): 22.44 seconds / 160.43 kernels/hour [P:9950%, 422 maj. pagefaults]
Run 2 (-j 512): 22.40 seconds / 160.71 kernels/hour [P:10000%, 135 maj. pagefaults]

Thanks Dell!

@gregkh I really like the idea of openbenchmarking.org / the #PhoronixTestSuite to compare multiple devices and sharing the results. The website is a bit clunky though, I wish more people were collaborating on this.

@gregkh
Processor: Blizzard-M2-Max [12 threads]
Compiler (target): x86_64-linux-gnu-gcc (Debian 15.2.0-4) 15.2.0
Linux compiled: 6.17.0 [/data/linux-6.17/]
Config; Environment: defconfig; ARCH="x86_64" CCACHE_DISABLE="1" CROSS_COMPILE="x86_64-linux-gnu-"
Run 1 (-j 12): 90.48 seconds / 39.79 kernels/hour [P:1023%, 54 maj. pagefaults]
Run 2 (-j 12): 90.39 seconds / 39.83 kernels/hour [P:1024%, 34 maj. pagefaults]

Laptop, x86-64 cross compile to keep it fair.

@axboe @gregkh but with what storage and memory?

@jeroen @gregkh If you're referencing one of mine, which one of them?

@gregkh fwiw the dual-screen laptop might be GPD, not GSD? at least i don't find one under S

thanks for sharing either way!!

@gregkh kernels/kWh I guess would be the way to take power into consideration.

@axboe @gregkh all, as while mentioning cores/ghz is fun, and then the amount of kernels it can build, the amount of mem (and speed of that) and the nvme/ssd/hdd even more change the numbers. And when mentioning numbers it can be good for those reading what components drive to a good amount of kernels builds (higher is less drinks fetched and consumed ;) -- though of course, what the config is matters too, it is about repeatability when debugging/dev

@jeroen @gregkh In my experience of compiling kernels, these days storage means very little unless it's really bad. The amd box I quoted has 32 NVMe drives in it, but ran on a single drive. If I changed that to tmpfs, I'd get basically the same numbers. Ditto on the laptop.

For kcbench, it uses the specific arch defconfig and it mentions compiler and other relevant details too.

@jeroen @gregkh My laptop run is a M2 Max 16" with 32G of RAM, as listed. The Dell box is a r7625 with 768GB of DDR5 at 4800 MT/s.

@jeroen @gregkh Case in point, here's the amd box running out of /dev/shm/ instead of on local storage:

Run 1 (-j 512): 22.46 seconds / 160.28 kernels/hour [P:9943%, 152 maj. pagefaults]
Run 2 (-j 512): 22.46 seconds / 160.28 kernels/hour [P:9947%, 146 maj. pagefaults]

@gregkh 4.70 kernels/hour on an RK3588 Radxa ROCK 5B running the stock Debian Trixie kernel 6.12.43, compiling kernel 6.17.0.

Not bad for something that costs USD 150 (or USD 110 for the 8GB version) and runs mainline.

@jeroen @gregkh and using clang-21:

Run 1 (-j 512): 28.82 seconds / 124.91 kernels/hour [P:14334%, 14705 maj. pagefaults]
Run 2 (-j 512): 28.86 seconds / 124.74 kernels/hour [P:14268%, 13373 maj. pagefaults]

because the compiler can make a massive difference.

@gregkh Great benchmark! Very interesting to see all your different machines. I did like what I saw of the Framework 13 at KR. 3:2 make so much more sense! I had been thinking earlier this year about getting the Framework 12 but it good to your opinion that it not quite fast enough.

@axboe @gregkh Yes, I agree; I tend to compile in tmpfs too because fortunately memory is plenty nowadays and given that and nvme along with caching disk speeds are mostly irrelevant unless one still has a hdd somewhere.

@jeroen @gregkh even in the hdd days storage didn't matter much for kcbench, unless it was extremely slow, as already pointed out by @axboe .

But they mattered a bit for the first run, which is why its results are thrown away by default. But I recently wondered if that is still necessary these days. 🤔

@kernellogger @jeroen @axboe When using "cloud" machines, many of them have horribly slow storage paths, so yes, it does affect kernel build times on those systems. Now if the "cache" really does anything for that type of storage or not, I do not know, but it can't hurt. Just like running sync 3 times before rebooting :)

@gregkh That was fun read! I'm missing RISC-V, and Alpine Linux machines though!

@ncopa You do NOT want to see the kcbench results for the riscv system I have here, it's so sad it's not even funny. So sad I haven't even powered it on in a few months, it's pretty much useless :(

@gregkh my 2012 dell laptop: 1 kernel build per hour XD

@gregkh @kernellogger Happy to read the last sentence, now I'm feeling less weird for running sync once after installing/updating software and before reboot.

@SchwarzeLocke @gregkh now I'm wondering if you have trust issues or if I'm too naive 🤔 🧐 🥴

@multisn8 Oops, yes, it is "GPD", I've now fixed the post, thanks for catching that.

@gregkh @ncopa on my risc-v system:

fn@rockos-eswin:~/tmp$ bash kcbench
[NOTE] Downloading source of Linux 6.8; this might take a while... 
Processor:           eswin,eic770x [4 threads]
Cpufreq; Memory:     ondemand [cpufreq-dt]; 9895 MiB
Linux running:       6.6.66-win2030 [riscv64]
Compiler:            gcc (Debian 14.2.0-19rockos1) 14.2.0
Linux compiled:      6.8.0 [/home/fn/.cache/kcbench/linux-6.8]
Config; Environment: defconfig; CCACHE_DISABLE="1"
Build command:       make vmlinux
Filling caches:      This might take a while... Done
Run 1 (-j 4):        1321.32 seconds / 2.72 kernels/hour [P:382%, 91 maj. pagefaults]
Run 2 (-j 4):        1320.99 seconds / 2.73 kernels/hour [P:382%, 84 maj. pagefaults]
Run 3 (-j 6):        1363.77 seconds / 2.64 kernels/hour [P:382%, 185 maj. pagefaults]
Run 4 (-j 6):        1369.81 seconds / 2.63 kernels/hour [P:381%, 221 maj. pagefaults]

@fn @ncopa Ouch, that's slow.