After benchmarking a Ryzen 7 255 mini PC, I got the machine I’d been weighing up: an ASUS ROG Flow Z13 with AMD’s Strix Halo chip and 128 GB of unified memory. This post compares it with the desktop I already had, an RTX 3090, on the same models and the same benchmark. It’s one of four posts from my local LLMs page, which also explains the stack and the model types if any of the terms here are new.
The short version: the 3090 is much faster on every model that fits in its 24 GB. The Z13 runs models the 3090 can’t.
The machines
| Laptop | Desktop | |
|---|---|---|
| Machine | ASUS ROG Flow Z13 | Desktop PC |
| CPU | Ryzen AI Max+ 395 (Strix Halo) | Ryzen 9 5900X |
| GPU | Radeon 8060S (integrated) | RTX 3090 |
| Memory for the model | 96 GB of the 128 GB, set aside for the GPU | 24 GB on the card |
| Backend | llama.cpp b12387, Vulkan | llama.cpp b12387, CUDA 12.4 |
| OS | Windows 11 | Windows 11 |
The laptop’s draw is memory: it can hold models far bigger than a 24 GB card can.
Speed
Same method as the mini PC post: llama-bench -ngl 99 -p 512,4096 -n 128 -r 3, every layer on the GPU, three repetitions. Prefill is how fast the model reads a prompt (512 and 4,096 tokens). Decode is how fast it writes. All numbers are tokens per second.
On the Z13:
| Model | Size | Prefill 512 | Prefill 4096 | Decode |
|---|---|---|---|---|
| Qwen3-30B-A3B Q4_K_M | 17.3 GiB | 1,373 | 1,174 | 83.3 |
| Nemotron 3.5 Lightning 30B-A3B Q4_0 | 17.6 GiB | 1,464 | 1,480 | 66.6 |
| gpt-oss-120b MXFP4 | 59.0 GiB | 830 | 804 | 52.8 |
| Qwen3.8-Flash-Next UD-IQ3_XXS | 77.9 GiB | 344 | 472 | 27.5 |
| Qwen3.8 27B Q4_K_M | 15.7 GiB | 287 | 265 | 13.0 |
On the RTX 3090 (gpt-oss-120b and Flash-Next don’t fit in 24 GB):
| Model | Prefill 512 | Prefill 4096 | Decode |
|---|---|---|---|
| Qwen3-30B-A3B Q4_K_M | 5,388 | 5,107 | 202.3 |
| Nemotron 3.5 Lightning 30B-A3B Q4_0 | 5,567 | 5,549 | 192.5 |
| Qwen3.8 27B Q4_K_M | 1,552 | 1,515 | 43.1 |
- The 3090 is much faster on every model that fits in its 24 GB: roughly 4–6× at prefill and 2.4–3.3× at decode.
- The Z13 runs models the 3090 can’t. Flash-Next and gpt-oss-120b don’t fit on the card at all.
- Qwen3.8 27B is a dense model, so it’s slow on the Z13. Every token reads all 27B weights. With its built-in multi-token prediction head (MTP) switched on in llama.cpp, decode went from 13 to roughly 24–30 tok/s in my coding runs.
What each machine is good for
Raw speed isn’t the whole story, because the fastest models weren’t the ones that did the work well. In my tests of writing animation scenes and unattended code repair:
- The 3090 was fastest when the model fit. Qwen3.8 27B with MTP and a capped thinking budget produced a 0.948 scene (out of 1) in under 10 minutes.
- The Z13’s best results came from a model only it can hold. Qwen3.8-Flash-Next scored 0.943 capped (21 minutes) and 0.972 uncapped (32 minutes) on the same scene, and passed all three repair projects.
- The fast mixture-of-experts models weren’t good enough. Qwen3-30B-A3B and Nemotron 3.5 Lightning decode several times faster, but never produced a working full scene on the 3090, and A3B failed two of the three repair tasks.
So the 3090 suits a model that fits in 24 GB and needs to go fast. The Z13 suits the bigger model you can queue up and leave running overnight. I also tried AMD’s new engine on the Z13: see Loom vs llama.cpp.
Caveats
- Windows only, so far. Linux, with ROCm, is next, and I’ll add those numbers to the local LLMs page.
- llama-bench is a synthetic test. It measures fixed prompt and output lengths. Real runs, where the model decides how much to write, vary more.
- Model settings differ between the scene and repair runs quoted above (MTP on or off, thinking capped or not, and a smaller context for Qwen3-30B-A3B on the 3090). Those posts have the details.