Arao AI
Back to search results

Qwen3-8B · IQ3_S

Start a discussion
NVIDIA H2097GBllama.cpp v0.28L2 Cross-framework verified

This page aggregates 1 real-world runs of Qwen3-8B (IQ3_S) on NVIDIA H20 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

57.35 tok/s

Decode

Decode speed

209.84 tok/s

Prefill

Prefill speed

0.38 s

TTFT

Time to first token

13.19 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

17 days ago

Last verified

Performance

  1. llama.cpp · IQ3_S (current)Decode 57.35 · Prefill 209.84 ·
  2. llama.cpp · Q5_K_S Decode 42.52 · Prefill 173.47 ·
  3. Strata · IQ3_S Decode 87.33 · Prefill 444.4 ·
  4. Strata · Q5_K_S Decode 59.66 · Prefill 317.37 ·

Core figures

Decode (avg)
57.35 tok/s
Prefill (avg)
209.84 tok/s
TTFT (avg)
0.38 s
VRAM (avg)
13.19 GB
MTP acceptance rate
—
TTFB
0.46 GB
Power draw
351.9 W

Configuration

Member-level fields are taken from the most recent run

Model
Qwen3-8B
Quantization
IQ3_S
Framework
llama.cpp
Version
v0.28
Context length
16384 tokens
Batch size
256
GPU layers
33
Flash Attention
Off

Hardware

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

GPU
NVIDIA H20
Nominal VRAM
97 GB
Measured VRAM (avg)
13.19 GB
OS
Windows 11 24H2
Driver
560.94
CUDA
12.8
Power draw
351.9 W

Sources & evidence

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

  1. L2 Cross-framework verifiedOtherOriginal link Verified on 2026-10-19

    v0.28 · Windows 11 24H2 · CUDA 12.8 · 16384 ctx

    57.35 tok/s

    Decode

    209.84 tok/s

    Prefill

    0.38 s

    TTFT

    13.19 GB

    VRAM

    —

    MTP

    351.9 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=56 running model_id=53 at IQ3_S