Arao AI
Back to search results

Qwen3-8B · FP16

Start a discussion
AMD Ryzen AI Max+ 395128GBllama.cpp v0.26L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-8B (FP16) on AMD Ryzen AI Max+ 395 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

52.16 tok/s

Decode

Decode speed

223.09 tok/s

Prefill

Prefill speed

1.39 s

TTFT

Time to first token

33.18 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

2 days ago

Last verified

Performance

  1. llama.cpp · FP16 (current)Decode 52.16 · Prefill 223.09 ·
  2. llama.cpp · Q8_0 Decode 70.95 · Prefill 365.06 ·
  3. vLLM · IQ3_S Decode 125.05 · Prefill 672.6 ·
  4. Ollama · IQ3_S Decode 106.51 · Prefill 487.62 ·

Core figures

Decode (avg)
52.16 tok/s
Prefill (avg)
223.09 tok/s
TTFT (avg)
1.39 s
VRAM (avg)
33.18 GB
MTP acceptance rate
—
TTFB
0.42 GB
Power draw
70 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-8B
Quantization
FP16
Framework
llama.cpp
Version
v0.26
Context length
16384 tokens
Batch size
512
GPU layers
20
Flash Attention
Off

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
AMD Ryzen AI Max+ 395
Nominal VRAM
128 GB
Measured VRAM (avg)
33.18 GB
OS
Ubuntu 22.04
Driver
560.94
CUDA
N/A
Power draw
70 W

Sources & evidence

1 measured records in total, each traceable to its original source

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-16

    v0.26 · Ubuntu 22.04 · CUDA N/A · 16384 ctx

    52.16 tok/s

    Decode

    223.09 tok/s

    Prefill

    1.39 s

    TTFT

    33.18 GB

    VRAM

    —

    MTP

    70 W

    Power draw

    L2 batch seed 2026-10-16: llama.cpp on hw_id=24 running model_id=53 at FP16