Datasets:
Ling-3.0-flash Q3_K_M: the --n-cpu-moe offload curve (RTX 5090)
Date: 2026-08-09 Author: WITCHEER Platform: NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)
What this measures
For MoE models too large for VRAM, llama.cpp's --n-cpu-moe N keeps the routed experts of the first N layers in system RAM and puts the rest on the GPU. The usual advice is "fill the card". This sweep measures what that actually buys on a 32GB RTX 5090 with Ling-3.0-flash Q3_K_M (58.3 GiB file, 127.5B total / 5.1B active, 512 routed experts), llama.cpp PR #26608 branch (head 0266ebca6), CUDA.
Method: llama-bench -ngl 99 --n-cpu-moe N -p 2048 -n 128 -r 3 per point, VRAM peak sampled at 1Hz via nvidia-smi during each run. Note the generation numbers are conditioned on a 2048-token prompt in context; bare tg128 on this model reads ~3 tok/s higher (46.0 in the Q3_K_M report).
Results
--n-cpu-moe |
gen tok/s | VRAM peak | note |
|---|---|---|---|
| 99 (all experts in RAM) | 42.9 | 3.6 GiB | baseline |
| 38 | 49.0 | 8.9 GiB | |
| 34 | 52.5 | 14.2 GiB | |
| 30 | 57.5 | 19.5 GiB | |
| 26 | 62.7 | 24.9 GiB | |
| 24 | 66.0 | 27.5 GiB | optimum, +54% over baseline |
| 22 | 63.6 | 30.2 GiB | inversion: more VRAM, less speed |
| 20 | — | OOM | allocation fails |
Reads
- Expert placement scales throughput almost linearly until the card is nearly full. Every ~5.5 GiB of experts moved to VRAM buys roughly 4-5 tok/s.
- The optimum is not at the VRAM ceiling. Peak throughput lands at 27.5 GiB on a 32GB card; pushing to 30.2 GiB costs 2.4 tok/s, and one step further fails to allocate. Leave ~4 GiB of headroom.
- The inversion at n_cpu_moe=22 is consistent with allocation pressure near the ceiling (fragmentation and reduced scratch/compute buffer room), not with any property of the model.
Raw JSON (per-point llama-bench output and VRAM samples) ships with the dataset. Companion quant-ladder reports: IQ3_XXS, IQ2_M.
