Hacker News Reader

Comment by usagisushi | Hacker News Reader

Not the OP, but their setup must be faster than my 4060 16GB + 3060 12GB setup. Here are my numbers (typical values, N=1):

    Model                         pp (t/s)    tg (t/s)
    Qwen 3.6 27B            900           29
    Qwen 3.6 35B-A3B   2100          85
    Gemma 4 31B            750           28
    Gemma 4 26B-A4B   2500         90

- All models: UD-Q4 w/ MTP. Context size: ~100k (MoE) / ~70k (Dense).

- Layer splitting used. Tensor splitting is ~1.2x faster in TG, but power spikes from 150W to 380W.

View on HN

In my case it somehow depends a lot on the task being performed... .

E.g. when doing text transcription/OCR from images (Qwen 3.6 27B Q4_K_M by Bartowski) with a context size of ~50k I get a pp of ~460 tokens per second and a generation ranging from 35 to 45 tokens per second (using "--spec-type draft-mtp --spec-draft-n-max 2" currently with llama.cpp b6548).

On the other hand when handling code (Qwen 3.6 27B Q5_K_M by Bartowski) with a context size of 128k I get a pp ranging between 500 to 1500 tokens per second and a generation between 25 and 40 tokens per second (using in this case as well "--spec-type draft-mtp --spec-draft-n-max 2" currently with llama.cpp b6548).

Anyway in theory with "--split-mode layer" I think that it's anyway the slowest card that drives the overall performance (I do see in "nvtop" that usually the 5070 is ~25% active, the 5060 ~50% and the 3060 ~75%).

Reply on news.ycombinator.com