It won’t be long until Linux enthusiasts will be able to complete a clean kernel build in under 10 seconds. Linux expert and Phoronix head honcho Michael Larabel said that “the coffee window is closing” in recent commentary that weighs computer processor advances and human / AI optimizations of the Kbuild code. What once used to be a computer processing task that provided a decent excuse for a coffee break now only allows enough time for a measured sip.
The source site’s determined Linux focus has meant that OS kernel compilation times have become a signature benchmark over the last two decades. Larabel notes that this once time-consuming process “can be done in now roughly 15 seconds.” It isn’t just advances in hardware and core counts pushing the envelope; compiling a defconfig x86-64 Linux kernel build has had many bottlenecks removed thanks to the work of Linux MM developer Lorenzo Stoakes and the assistance of AI/LLMs recently. Even modest processors have seen their compile times cut dramatically thanks to this work.
Larabel went hands-on with the latest v4 kernel patches from Lorenzo, and armed with his very powerful EPYC workstation, achieved “a 15-second kernel build!” Before the latest patches, the same system took over 22 seconds for the same compilation task.
The test system is extremely potent, though. Your home PC or laptop might still let you have enough time to go brew some coffee, as Larabel’s test rig features the following components:
Samsung PM1743 3.84TB PCIe Gen 5.0 NVMe SSD storage
Stock Ubuntu 26.04 LTS
It is noted that the code compilation benchmarking didn’t employ a RAMdisk to aggressively tune the compile times
All the above timings mentioned are for the minimal default kernel compilation for x86-64 systems. The Phoronix chief also timed the compilation of a Linux kernel with ‘allmodconfig’ with thousands of drivers and subsystems enabled. This precipitates far longer compile times. On the above beefy system, though, the latest Lorenzo patches drove the time down from 169 seconds to 134 seconds. So, this other milestone is approaching the two-minute mark.
Phoronix is now taking bets on AMD (Zen 6) EPYC 9686F systems, with MRDIMMs and PCIe Gen6 storage pushing defconfig x86-64 Linux kernel compilation times under the 10-second mark.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Mark Tyson is a news editor at Tom's Hardware. He enjoys covering the full breadth of PC tech; from business and semiconductor design to products approaching the edge of reason.
There are a lot of caveats, like the test system having 128 cores, but Zen 6 and Nova Lake are going to bring big multi-threaded increases to consumer sockets. And if you're sitting the generation out because of high DDR5 prices, Zen 7 should bring another +33% to core counts, and Hammer Lake should bring back SMT.
Wait a second!
Are you saying 256 threads, 576 Gigabytes a second of memory bandwidth and 14 gigabytes a second of SSD bandwidth allows you to compile faster?
Wait a second!
Are you saying 256 threads, 576 Gigabytes a second of memory bandwidth and 14 gigabytes a second of SSD bandwidth allows you to compile faster?
The news is that many of the single-threaded bottlenecks have been removed from the Linux kernel build chain so it scales better to multi-core systems. For other software projects there may be configure scripts and build dependencies that prevent effective parallel scaling. In those cases, faster cores win over more cores.
It would be interesting to report the average percent CPU use for a kernel build running on a 128 core workstation. For a less anecdotal story one could measure average percent CPU use versus number N of available cores for values of N ranging from 1 to 128.
I know build system like `make` launches as many parallel processes as it can/is allowed - but do modern compilers and linkers also use multiple thread or they are still single threaded?
It should be mentioned that most builds a software developer does are incremental, and thus only require a small subset of the source files to be recompiled. Tools like ccache can be used to further reduce compilation times, such as when switching between branches.
There are a lot of caveats, like the test system having 128 cores, but Zen 6 and Nova Lake are going to bring big multi-threaded increases to consumer sockets.
I wouldn't put them in the same category. Compilation is very bandwidth-intensive. In this article, Phoronix tested the performance impact of running a 270K+ Arrow Lake on 1 DIMM vs. 2.
https://www.phoronix.com/review/single-dual-memory-linux
With two DIMMs, kernel compilation was 13.8% faster. Since the compute-to-bandwidth ratio for a 52-core Nova Lake should be similar as a 270K+ running on a single DIMM, I wouldn't expect it to beat even a 48-core Threadripper with quad-channel memory.
Wait a second!
Are you saying 256 threads, 576 Gigabytes a second of memory bandwidth and 14 gigabytes a second of SSD bandwidth allows you to compile faster?
What's funny about that question is that Linux had some notable scaling problems, when it came to building the kernel on systems with more than ~64 cores.
Yes, this is news. Not consequential for the general public, but for people developing & deploying software on big 100+ -core servers/VMs, it's noteworthy.
I know build system like `make` launches as many parallel processes as it can/is allowed - but do modern compilers and linkers also use multiple thread or they are still single threaded?
Compilers like GCC and Clang still run single-threaded per source file. When you have enough files in your codebase, relative to the number of cores in a machine, it's more efficient for the compiler to stay single-threaded.
In modern development toolchains, the main place you see multi-threading is in linkers. Those tend to happen near/at the end of a build and can take a decent chunk of time. So, it makes a lot more sense to focus on concurrency there.
With two DIMMs, kernel compilation was 13.8% faster. Since the compute-to-bandwidth ratio for a 52-core Nova Lake should be similar as a 270K+ running on a single DIMM, I wouldn't expect it to beat even a 48-core Threadripper with quad-channel memory.
We'll see. I think Nova Lake should be supporting faster memory than Arrow Lake (e.g. native 8000 MT/s, 10000 MT/s overclocked), and faster memory than Threadripper users are running.
