Arao AI
Back to search results

Llama-3.2-3B-Instruct · Q3_K_S

Start a discussion
NVIDIA H2097GBllama.cpp v0.51L2 Cross-framework verified

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (Q3_K_S) on NVIDIA H20 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.

171.35 tok/s

Decode

Decode speed

761.41 tok/s

Prefill

Prefill speed

2.49 s

TTFT

Time to first token

6.2 GB

VRAM

VRAM usage

L2 Cross-framework verified

1

Measured runs

1

Independent sources

Other

Source platforms

16 days ago

Last verified

Performance

  1. llama.cpp · Q3_K_S (current)Decode 171.35 · Prefill 761.41 ·
  2. vLLM · FP16 Decode 76.18 · Prefill 254.17 ·
  3. vLLM · IQ3_XS Decode 177.1 · Prefill 886.15 ·

Core figures

Decode (avg)
171.35 tok/s
Prefill (avg)
761.41 tok/s
TTFT (avg)
2.49 s
VRAM (avg)
6.2 GB
MTP acceptance rate
—
TTFB
1.37 GB
Power draw
353.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
Q3_K_S
Framework
llama.cpp
Version
v0.51
Context length
8192 tokens
Batch size
256
GPU layers
99
Flash Attention
On

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)
6.2 GB
OS
Windows 11 23H2
Driver
572.83
CUDA
12.4
Power draw
353.7 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

    v0.51 · Windows 11 23H2 · CUDA 12.4 · 8192 ctx

    171.35 tok/s

    Decode

    761.41 tok/s

    Prefill

    2.49 s

    TTFT

    6.2 GB

    VRAM

    —

    MTP

    353.7 W

    Power draw

    L2 batch seed 2026-10-19: llama.cpp on hw_id=56 running model_id=70 at Q3_K_S