When Phoronix pitted identical AMD hardware against itself in a cross-platform AI face-off, the results surprised no one who follows open-source driver development: Linux, armed with the latest community-built graphics drivers, outran Windows 11 in Llama.cpp’s Vulkan-accelerated inference tests. The tests, published in late March 2025, used a Ryzen 9 9950X3D CPU and Radeon RX 9070 XT GPU, and the performance gap underscores how rapid iterations in the open-source Mesa RADV driver and the Linux kernel can tip the scales for local AI workloads.

What the Benchmarks Actually Found

Phoronix ran native builds of Llama.cpp on the same machine, booting first into Windows 11 25H2 with AMD’s Radeon Software 25.8.1 driver, and then into Ubuntu 24.04.3 LTS with two different Linux configurations. The baseline Linux setup used the HWE (Hardware Enablement) stack with kernel 6.14 and Mesa 25.0; the second, more revealing configuration swapped in a development Linux kernel 6.17 and Mesa 25.3-devel fresh from the RADV repository. The GPU was a Radeon RX 9070 XT, a mid-range RDNA4 card with 16 GB of VRAM.

The headline: on Vulkan-accelerated AI inference with Llama.cpp, the upgraded Linux stack turned in noticeably higher token generation rates than Windows. The advantage wasn’t uniform across every model or workload profile, but it was especially pronounced in scenarios where Vulkan dispatch overhead and memory handling dominate—exactly the areas that the Mesa RADV team has been systematically optimizing for RDNA4.

CPU-only inference also saw a Linux boost thanks to newer kernels and compiler toolchains, a well-established pattern for multi-threaded workloads on Zen-class processors. But the GPU side is what demands attention: it means the same silicon can spit out tokens faster depending solely on which operating system and driver stack you choose.

Exact percentage gains aren’t fixed. Phoronix’s charts show clear per-test deltas, but those numbers are tied to the specific driver and kernel commits used. Reproducing the test with a different Mesa snapshot could shift the magnitude. The takeaway isn’t a single digit; it’s that Linux’s open-source driver ecosystem now delivers a real, repeatable edge for local AI inference on RDNA4 hardware.

Why It Matters for Your Setup

The performance gap changes the calculus for different kinds of Windows users.

For home users and hobbyists who run large language models locally, the numbers are a quiet call to action. If you’re already comfortable dual-booting or spinning up a Linux virtual machine with GPU passthrough, you can wring more speed out of your existing RDNA4 GPU at no extra cost. Llama.cpp works identically on both platforms, so moving to Linux for inference is mostly a driver question.

For power users and workstation builders, the message is sharper. When you’re tuning a machine specifically for AI tinkering—where every token per second counts during long generative sessions—the choice of host OS becomes a performance dimension. A dedicated Linux partition with a rolling Mesa build could be the difference between a model feeling responsive and feeling sluggish.

For IT professionals and DevOps teams evaluating inference infrastructure, the benchmarks provide evidence for hybrid deployment models. A Linux GPU node running the latest RADV driver might serve inference APIs faster than an equivalent Windows node, freeing the Windows boxes for tasks that require vendor-certified driver stacks or Windows-only software.

None of this is to say Windows is broken for AI. It works, and AMD’s official drivers remain the safe, supported path. But the benchmarks make it clear that for this specific workload—Vulkan-accelerated Llama.cpp inference on RDNA4—the open-source stack has pulled ahead.

The Open-Source Advantage: A Deeper Look

How did a free, community-built driver overtake AMD’s own Windows driver? The answer lies in the sheer pace of iteration and the types of improvements that landed in Mesa 25.x throughout early 2025.

Mesa’s RADV driver, which implements Vulkan on AMD GPUs for Linux, is perpetually under active development by a mix of Valve engineers, independent contributors, and AMD employees. Because it’s open source, new performance work and fledgling Vulkan extensions get merged, tested, and shipped in weekly or bi-weekly snapshots. For an AI runtime like Llama.cpp, three categories of RADV improvements matter enormously:

  • Bfloat16 support. The VK_KHR_shader_bfloat16 extension and related bfloat16 pathways graduated into Mesa’s development branches during the 25.x cycle. Lower-precision arithmetic reduces VRAM bandwidth pressure and allows AI kernels to execute more compute per byte, directly lifting token throughput.
  • Micro-optimizations and low‑level fixes. RADV has a history of small patches producing outsized gains. Valve’s culling work in previous releases is legendary, and equivalent attention has been paid to the compute dispatch path that Llama.cpp rides. When a driver trims even a few dozen CPU cycles from each Vulkan command submission, the cumulative effect over thousands of tokens is measurable.
  • RDNA4-specific plumbing. The RX 9070 XT’s GFX12 architecture required fresh RADV code. The team invested in BVH infrastructure, queue management, and device memory handling that, while necessary for ray tracing and gaming, also smooth out the compute pipeline AI workloads lean on.

The Linux kernel plays a supporting role. Kernel 6.16–6.17 development cycles brought amdgpu power-management patches and scheduling tweaks that lower CPU overhead for GPU submissions. Together, a bleeding-edge kernel and Mesa RADV form a pipeline that wastes fewer cycles between the moment Llama.cpp asks for a compute dispatch and the moment the GPU executes it. On Windows, the official driver—while stable and certified—hasn’t historically chased these incremental compute-focused optimizations with the same urgency.

How to Get the Best Performance Today

If you want to chase the Linux advantage, here’s what to do right now.

Keep your Windows drivers current. AMD regularly updates Radeon Software with performance improvements. Even if the open-source stack is faster today, a future Windows driver drop could narrow or reverse the gap. Check for updates in the AMD Adrenalin software or on AMD’s website.

Experiment with Linux on a spare partition or external drive. Installed alongside Windows, a rolling distribution like Arch Linux or the latest Ubuntu interim release with the oibaf PPA for Mesa can give you a quick testbed. Boot Linux, install Llama.cpp, and run your own before‑and‑after benchmarks using identical models and prompts. The Phoronix results are a starting point, not a guarantee—your model mix and usage pattern may see different gains.

Pin your software versions when you measure. The difference between Mesa 25.3-devel commit abc123 and commit def456 could be a few percent. If you’re going to tune a production system, record the exact kernel and Mesa hashes so you can reproduce your results later.

Consider a hybrid setup for serious work. Spin up a dedicated Linux machine or a GPU‑passed VM as an inference worker. Call it from your Windows daily driver via a local API. This isolates the experimental driver risk to a single node while letting you keep the stability of Windows for everything else.

Don’t chase every nightly. Bleeding‑edge Mesa and kernel release candidates can introduce regressions. The performance edge is real, but so is the chance of a crash mid‑inference. Test in a non‑critical environment before deploying into any workflow you depend on.

What’s Next for Windows Users

AMD’s Windows driver team isn’t standing still. The upcoming Radeon Software releases will almost certainly incorporate some of the same Vulkan optimizations that have already landed in RADV, especially as AI workloads become a higher priority. The performance landscape is fluid: what’s true in March 2025 may shift by summer.

For now, the open‑source stack gives Linux a measurable leg up on RDNA4 GPUs for Llama.cpp’s Vulkan inference. If you’re building a machine primarily for local AI, it’s worth adding a Linux option to your plan. If you’re sticking with Windows out of habit or necessity, keep your drivers fresh and watch the benchmarks—the pendulum may swing back.