Across 99 comparison benchmarks running the high-end laptop "Razer Blade 18" on Windows 11, Ubuntu 26.04, and CachyOS, Windows showed strength in certain local AI processing scenarios. The gap reportedly reached as much as 3.43 times that of Ubuntu at its peak. However, tracing through the raw data and test profiles reveals that this figure doesn't directly indicate OS superiority. In the llama.cpp CPU test, Windows and Linux followed different build paths, while Vulkan text generation came out nearly tied. Meanwhile, Windows' advantage persisted in speech recognition, which used identical distributed binaries. These two results need to be read separately.
CachyOS Takes First Place Across 99 Tests, Windows Often Ranks Last
The comparison was conducted by Phoronix, using a Razer Blade 18 equipped with a Core Ultra 9 290HX Plus, GeForce RTX 5090 Laptop GPU, and 32GB of DDR5-6400. Windows 11 was the environment as shipped by Razer, Ubuntu used version 26.04 LTS with Linux 7.0, and CachyOS used the rolling release with Linux 7.1.
Even on the same chassis, software conditions weren't uniform. The NVIDIA driver was 596.36 on Windows, 595.71.05 on Ubuntu, and 610.43.02 on CachyOS. The Linux side also differed in kernel, compiler, and file system. This is a test comparing each OS's realistic configuration, not an experiment holding everything constant except the OS.
Across the 99 tests, CachyOS took first place 51% of the time, Windows 38%, and Ubuntu 10%. Conversely, Windows finished last 50% of the time, and Ubuntu 43%. In the geometric mean across all results, CachyOS held a slight lead, and Windows edged out Ubuntu by a small margin. Linux showed strength in rendering tasks like Blender and in video encoding with OpenAPV and SVT-AV1. You cannot extrapolate from a subset of AI tests that "Windows overwhelmingly dominates Linux across the board."
The Peak 3.43x Gap Comes from the CPU Path; Vulkan Generation Is Nearly Even
The gap in local AI performance varies significantly more by execution path and processing stage than by model name. Lining up representative values makes the outline clear.
| Test | Metric | Windows 11 | Ubuntu 26.04 | CachyOS | Windows vs Ubuntu |
|---|---|---|---|---|---|
| llama.cpp / GLM / CPU | Prompt Processing | 116.62 tok/s | 33.97 tok/s | 74.21 tok/s | 3.43x |
| llama.cpp / gpt-oss / Vulkan | Prompt Processing | 6027.71 tok/s | 3801.76 tok/s | 3832.61 tok/s | 1.59x |
| llama.cpp / gpt-oss / Vulkan | Text Generation | 184.55 tok/s | 184.36 tok/s | 186.96 tok/s | Nearly Identical |
| Whisperfile / medium | Speech Recognition Time | 520.87s | 991.70s | 976.80s | 1.90x faster |
The 3.43x figure comes from running llama.cpp's GLM model on CPU, processing a 2048-token prompt. In text generation with 128 tokens, the CPU-path Windows outperforms Ubuntu by 1.59 to 2.81 times across several items.
The GPU Vulkan path told a different story. CachyOS took first place in all three text generation tests—Qwen, gpt-oss, and GLM—but the gap with Windows was at most 1.31%. Yet in gpt-oss prompt processing, Windows was 1.59 times faster than Ubuntu, and 1.08 times faster with GLM. Even with the same API, the process of reading input and the process of generating the next token can't be bundled into a single number.
Same "CPU BLAS" Label, Different Underlying Build
The key to interpreting the maximum gap lies in OpenBenchmarking's llama-cpp-2.5.0 profile. On Linux, the llama.cpp b8933 source is extracted, and it's compiled on the spot with GGML_BLAS=ON and GGML_BLAS_VENDOR=OpenBLAS specified. On Windows, rather than building the same revision's source with the same settings, the officially distributed llama-b8933-bin-win-cpu-x64.zip is extracted instead.
Looking at llama.cpp's b8933 release process, the Windows CPU package is built with GGML_CPU_ALL_VARIANTS=ON, GGML_BACKEND_DL=ON, and GGML_OPENMP=ON. The actual distributed binary includes CPU DLLs supporting Alder Lake, Ice Lake, Sapphire Rapids, Zen 4, and others, and the relevant job has neither GGML_BLAS=ON nor an OpenBLAS vendor specification. Even though the test name includes "BLAS," this isn't a comparison where both OSes use the identical compute backend.
Thread counts matched to revision, model, and physical core count are consistent across both. Still, when compiler, optimizations, and CPU kernel selection differ, you can't isolate the OS's contribution alone from within the 3.43x breakdown. What this figure shows is only "how the executable prepared for each platform performed on this one machine." It's not proof that Windows' kernel or scheduler is 3.43 times better than Linux's.
The 1.90x from Whisperfile Remains a Separate Signal
Even after identifying the build differences, it would be premature to dismiss Windows' advantage entirely. The Whisperfile profile uses the same .llamafile distributed by Mozilla and the same audio input across each OS. With the tiny model, Windows took 45.41 seconds versus Ubuntu's 66.40 seconds—a 1.46x gap. With small, it was 192.30 seconds versus 342.17 seconds—1.78x. With medium, it was 520.87 seconds versus 991.70 seconds—a 1.90x gap. CachyOS stayed close to Ubuntu across all sizes, while Windows was faster throughout all three sizes.
Since the distributed binary and input are identical here, the conditions across platforms are closer than in the llama.cpp CPU comparison. At minimum, the fact remains that Windows held an advantage in this specific combination of Razer Blade 18 and Whisperfile. However, the profile was created in August 2024, and only a single audio input was used. There's no basis for extrapolating this 1.90x figure to current Whisper-family implementations in general, to other CPUs or desktop PCs, or to GPU inference.
Ubuntu Certification and Razer's Linux Support Aren't the Same Thing
According to Phoronix, this Blade 18 is currently in the process of becoming the first Razer laptop to receive Ubuntu certification. Canonical's desktop certification is a system that verifies hardware compatibility with specific Ubuntu releases and monitors for regressions introduced by updates. It doesn't guarantee that individual applications will run faster than on Windows.
Razer's official support draws an even clearer line: while installing a different OS is possible, the company states that its products are designed for the OS they ship with, and both warranty and support are limited to that shipped OS. Even once Ubuntu certification is complete, it will carry meaningful weight for compatibility verification on the certified release, but it won't automatically change Razer's own support policy.
To properly measure OS-level performance differences, you'd need to align not just the llama.cpp revision but also the CPU backend and compiler settings, while fixing thread counts. Power modes and thermal conditions would also need to match. For Vulkan, driver generations should be aligned, with median values and variance disclosed. For Whisperfile, it would be worth confirming reproducibility with current builds and multiple audio samples. What the original data tells us is that Windows can become dramatically faster along certain local AI execution paths. It's not a verdict on Linux as a whole.