Ryzen 9 9950X Performs 16% Faster On Intel-Optimized Linux Distro

Phoronix's Michael Larabel benchmarked AMD's latest Ryzen 9 9950X in several different Linux distros and found that the Zen 5 chip performs up to 16% faster with the Intel-optimized Clear Linux distro. Here's an excerpt from the report: The Linux distributions for this round of testing on the AMD Ryzen 9 9950X included Arch Linux, CachyOS, Clear Linux, Fedora Workstation 40, Ubuntu 24.04 LTS, and a recent daily snapshot of Ubuntu 24.10 in its current development form. Intel's Clear Linux is the one most interesting for looking at on the new AMD Zen 5 hardware. While there hasn't been so much Clear Linux news in recent times, it remains the most well optimized x86_64 Linux distribution out of the box. Clear Linux makes use of compiler function multi versioning, performance-minded defaults, aggressive compiler CFLAGS/CXXFLAGS defaults, optional AVX-512 usage for more libraries, and many other patches and optimizations in the name of delivering the greatest x86_64 Linux performance. And while not Intel's focus, it works typically on AMD hardware too. [...]

Using the same Ryzen 9 9950X system, all of these Linux distributions were tested in their default / out-of-the-box state. [...] When taking the geometric mean of 59 benchmarks run across all of the Linux distributions on this AMD Ryzen 9 9950X system, Intel's Clear Linux easily took the crown. Ubuntu 24.04 LTS -- which was used for all of the Ryzen 9000 series Linux testing so far on Phoronix -- was the slowest. Tapping Intel's Clear Linux netted a 16% improvement on top of the performance offered by Ubuntu 24.04 LTS! Ubuntu 24.04 with the Ryzen 9000 series was already looking great generationally, but as shown today the performance can be even better with further software optimizations.

The Arch Linux powered CachyOS that is tuned out-of-the-box with a similar aim to Clear Linux also performed great. CachyOS was 7% faster than Ubuntu 24.04 LTS based on the geo mean and 3% faster than upstream Arch Linux itself. For different workloads though the CachyOS advantage over Arch Linux varied from a minimal difference to quite significant advantages. From the performance of PHP and Python scripts atop Clear Linux to compiling various server and HPC minded software, Intel's Clear Linux -- and a commendable second place for CachyOS -- were showing that even greater performance can be achieved on the AMD Ryzen 9 9950X. Even for devoted Ubuntu Linux users, these results did show some nice advantages of the upcoming Ubuntu 24.10 release over Ubuntu 24.04 LTS thanks to the GCC 14 compiler. Ubuntu 24.10 performance is also still subject to change since the current daily ISOs haven't yet moved past the Linux 6.8 kernel while Ubuntu 24.10 in October will be shipping with Linux 6.11.

Gentoo

[Bert64]

How about Gentoo if you build it using appropriate flags specifically for Ryzen?

Re: Gentoo

[reanjr]

It'd probably be close, but good luck beating Intel's engineers at tweaking compiler flags.

I'm more interested to see what the performance is like if they used Intel's compilers. It's not really a surprise that Intel's distro beats the competition; what's surprising is they don't use their own compiler.

Re:

[Bert64]

Not everything can be compiled with the intel compiler, and sometimes it's necessary to compile all dependencies with the same compiler or you can end up with breakage. In practice you can only really use a non-gcc compiler for building some specific things.

Re:

[short]

OpenMandriva claims to be fully clang-built: https://en.wikipedia.org/wiki/... [wikipedia.org] Also what you say would mean the compiled is not ABI compliant. At least clang tries to be gcc ABI compliant.

Re:

[drinkypoo]

icc hasn't been faster than gcc for years. It was the best thing around for a long time, but that time is gone.

Re:

[AmiMoJo]

Maybe Intel did optimize their CPU. If it dies after a couple of years and you are forced to buy a new one, that's great news for Intel. The excessive voltage probably allowed them to benchmark a bit more competitively too.

Their real issue is that their architecture is old and needs a major refresh like Ryzen was for AMD, plus their manufacturing process isn't competitive.

stable?

[korgitser]

Agressive cflags are nice and all, but no ricer worth their salt will ever admit to the stability issues you can run into. That Phoronix said no word about this does not surprise me.

Without knowing anything about any such effort behind it, in general I would totally expect the lts distro to be the most conservative in that regard. People use it expecting it to be stable.

Re:

[bill_mcgonigle]

> the stability issues you can run into

Especially with tricks like loop unrolling the compiler may guess that your code does something other than what it really does.

Writing plain boring code that doesn't rely on side-effects helps avoid this problem.

I know, we were trying to save every opcode in the 80's. Now we want the compiler to be able to optimize.

Different tricks for different eras.

Re:

[TeknoHog]

Writing plain boring code that doesn't rely on side-effects helps avoid this problem.

I used to have -mfpmath=sse,387 in my Gentoo CFLAGS. Then late last year after a Glibc update, the Chia cryptocurrency daemon started acting up. The daemon was a binary install, not compiled on my system. I didn't see any issues with any other program, until one day in the Python interpreter I noticed some floating point results looked off. For example log(exp(1)) returned 1.0018138846048894. Somehow as Glibc was updated from 2.38 to 2.39 or so, it became more sensitive to this flag and something broke.

O

Re:stable?

[thegarbz]

Clear Linux is not about pushing agressive cflags, it's about optimising scheduling, cpu govenors, power subsystems, and generally the use of hardware for specific purposes. But you're really begging the question. Maybe Phoronix said no word about stability because there's nothing to say about it. Just a quick Google search on Clear Linux shows no major topics concerning stability, just the usual complains about all distros being "it doesn't use package manager X which I think is better than Y" and stuff like that.

Given that as you say people who use it expect it to be stable I'd expect to actually find information online to the contrary. After all people love to complain more than anything.

Re:

[PhrostyMcByte]

"Aggressive" is a loaded term here.

Optimizations don't alter behavior of code. Stability issues are not the fault of the compiler or cflags used, but rather an issue in the code being compiled.

Don't depend on undefined behavior like aliasing visibility. Use warnings.

The only bad optimization I can think of is fast-math, which will impact code that depends on carefully crafted floating point ops to maintain high precision. My assumption is that they have not enabled this, except for apps which don't depend o

Re:

[short]

-ffast-math also needs -fno-finite-math-only otherwise it breaks code a lot.

Re:

[drinkypoo]

Optimizations don't alter behavior of code.

Just cranking up -O can break some code. I don't know why, that's not my department, I've just learned not to try to use more than -O3 without a good reason.

Re:

[PhrostyMcByte]

Endeavour to understand the tools you're using! Don't leave it up to mysticism.

Is AMD the place to go for AVX-512 at this point?

[fuzzyfuzzyfungus]

Intel may not care; since their Xeon systems still support it; but(ironically) the fact that their consumer tier chips are heterogeneous architecture(with the smaller cores not supporting AVX-512; and the instructions disabled across the board when the smaller ones are enabled to avoid the scheduler challenges of having certain processes that are only supported on some cores); and AMD's either being homogenous or heterogenous with their 'c' cores that are identical in instruction set; it may have become the

Re:

[Rockoon]

Intel isnt even the place to go for AVX-512

They should have stopped at 256 bit registers and extended the simd instruction set to work directly with memory of arbitrary size, or extended it to support some useful horizontal operations as well as 128-bit integer

LMAO!!!

[erasmix]

Intel’s going the way of Boeing! These POS companies can’t make anything anymore! I can’t say I’m sorry for them.

Not suprising

[wakeboarder]

AMD has to be able to run intel optimized code well, because a lot of code is optimized for intel