With the AMD Zen 5 generation, the timing is interesting where it’s not the desktop processors launching first but happens to be in the form of AMD Ryzen AI 300 series laptops. With the last minute delay of the Ryzen 900 series by 1~2 weeks, the embargo lift for the Ryzen AI 300 series is timed for this Sunday morning where I can now present the first AMD Zen 5 Linux benchmark results. And with being the first Zen 5 chip in my lab, I have been pushing it hard… Here is an extensive look at the ASUS Zenbook S 16 I received with the Ryzen AI 9 HX 370 current flagship SoC compared to a variety of other AMD and Intel laptop models. The focus was on both the raw performance and the package performance-per-Watt for the overall power efficiency of this Zen 5 SoC. And with it being the first Zen 5 hardware in the lab, I didn’t limit the selection to just conventional laptop workloads but also explored the performance characteristics for various other workloads of interest to diverse Linux users and for an idea of the HX 370 potential or similar Zen 5 chips appearing in thin client / edge / IoT type devices. This initial taste of AMD Zen 5 has me extremely excited about the performance potential of the upcoming Ryzen 9000 series and EPYC Turin processors.
AMD kindly supplied an ASUS Zenbook S 16 with Ryzen AI 9 HX 370 SoC review sample to provide for launch-day Linux testing at Phoronix. As back during the Computex announcement I pre-ordered an ASUS Zenbook S 16 with Ryzen AI 9 365, the review and comparison with that nearly identical laptop but slightly different SoC will be coming up in the days ahead on Phoronix. In fact, that AMD Ryzen AI 9 365 laptop already arrived early at Phoronix yesterday (the other ASUS Zenbook in these pictures) and is undergoing Windows vs. Linux benchmarks and plenty of other complementary tests between the Ryzen AI 9 365 and Ryzen AI 9 HX 370.
The ASUS Zenbook supplied by AMD only arrived last Tuesday and thus with less than one week of testing this article is my very early look at its performance and support. So expect more follow-ups in the coming days and weeks. For the most part my Linux expectations for the AMD Ryzen AI 300 series were in order. The AMD Zen 5 CPU support appears in good standing: a relatively recent kernel, GCC 14.1+ if caring about Znver5 targeting, etc. The main caveats are around the GPU and Ryzen AI NPU.
For the GPU you will need Linux 6.10+ and Mesa 24.2+ for the best support of the RDNA3.5 integrated graphics. So if installing Ubuntu 24.04 LTS out-of-the-box, you will be running LLVMpipe and without any accelerated graphics. But the kernel and Mesa upgrade are straight-forward. There is the necessary firmware as well within linux-firmware.git already. But in my testing on the ASUS Zenbook S 16 I began encountering kernel errors and ultimately a system hang around 20 minutes into the boot process. AMD Linux engineers ended up providing me with newer PSP and DMCUB firmware that hadn’t yet been upstreamed to linux-firmware.git. With those updated firmware files, the ASUS Zenbook laptop with Ryzen AI 9 HX 370 has been running stable and not experiencing the hangs atop Linux 6.10 + Mesa Git. Those updated firmware files will be presumably made public in linux-firmware.git soon — stay tuned to Phoronix news to hear when that occurs. For more on the RDNA3.5 GPU performance, see today’s AMD Radeon 890M “RDNA3.5” Graphics Run Well With Latest Open-Source Linux Driver article.
On the NPU side, the XDNA Linux driver is finally going through its upstreaming process for getting into the mainline Linux kernel. Given the release cycle timing and distribution timings as well as waiting to see more user-space software make use of the Ryzen AI NPU stack, that’s now more something to look forward to in the very late 2024 or spring 2025 timeline for those running out-of-the-box on the likes of Ubuntu. The benchmarks in this article are all focused on the CPU side.
One other issue I encountered in my initial testing was the RAPL/PowerCap CPU package power information not being reported under Linux 6.10… Unfortunately, the Family 1Ah check wasn’t added to the RAPL driver in time. There is a patch adding that one-liner to allow for CPU power reporting with Zen 5 processors. Hopefully that patch will be upstreamed soon to the kernel and back-ported to stable series. For my ASUS Zenbook testing I used a Linux 6.10 build with in effect the same patch carried.
Another item to be aware of is the heterogeneous core topology handling. AMD engineers have been working on AMD P-State patches for heterogeneous core topology CPUs but as of writing the patches haven’t yet been upstreamed into the mainline Linux kernel. These patches should help with better placement of processes onto the Zen 5 cores rather than Zen 5C. But as these patches weren’t picked up for the Linux 6.11 merge window, unless they end up being submitted as a “fix” for Linux 6.11, that means they won’t be merged until at least the Linux 6.12 kernel cycle. Linux 6.12’s merge window will be in September while that stable kernel won’t be out until near the end of 2024. So unless patching your own kernel or if these AMD P-State patches get picked up as a “fix” for v6.11, you may be waiting a while for better Zen 5 vs. 5C handling. With the Ryzen AI 9 HX 370 there are four Zen 5 cores and eight Zen 5C cores.
In preparing for the AMD Ryzen AI 300 series Linux testing, I spent the past several weeks re-testing various laptops available in my lab. I was carrying out a wide range of benchmarks in being curious about the performance but with even more added value now that the Ryzen 9000 series desktop processors aren’t launching until 8~15 July. So in being eager to see how AMD Zen 5 is performing over Zen 4 and competitor Intel cores, there is a lot of workloads that typically aren’t run on laptops being shown in this article. It’s just a teaser of what to expect with Ryzen 9000 series and EPYC Turin later in the year if all the generational uplift holds true. Servers, HPC, technical computing, and other areas should especially be a real treat with Zen 5 — keep in mind with the Strix Point laptop tested is still with a 256-bit data path and not even the 512-bit path new to Zen 5 for desktop and server.
Originally before learning of the increased lead time between Ryzen AI 300 and Ryzen 9000 series availability, my other motivation for the diverse loads with this laptop testing was given the number of vendors offering Ryzen SoCs re-targeted for thin clients / embedded / mini computers / IoT type workloads that may run more database and server type workloads, more applicable to Linux software stacks, etc. So while testing in the laptop form factor today, the results excite me for what we may see of the AMD Ryzen AI 300 series possibilities (or under the Ryzen Embedded branding) in other form factors.
Page 1 – IntroductionPage 2 – Linux Laptop TestingPage 3 – Code Compilation BenchmarksPage 4 – 7-Zip, Blender, V-RAY, OSPRay BenchmarksPage 5 – OSPray Studio + Coremark + Embree BenchmarksPage 6 – QuantLib + OpenRadioss + HPC BenchmarksPage 7 – NAMD + Apache + Cryptsetup BenchmarksPage 8 – Crypto + Databases + Memcached BenchmarksPage 9 – Python + Perl Scripting, Simdjson AVX-512 BenchmarksPage 10 – Video Encoding BenchmarksPage 11 – AI Benchmarks + MorePage 12 – Octave + Darktable + Stockfish + Y-Cruncher + Web Browser BenchmarksPage 13 – Geo Mean + Power Use
