Arao AI
Back to search results

Llama-3.2-3B-Instruct · FP16

Start a discussion
NVIDIA H2097GBvLLM v0.47L2 Cross-framework verified

This page aggregates 1 real-world runs of Llama-3.2-3B-Instruct (FP16) on NVIDIA H20 with vLLM, contributed by 1 independent source platforms; metrics are averages of published measurements.

76.18 tok/s

Decode

Decode speed

254.17 tok/s

Prefill

Prefill speed

1.99 s

TTFT

Time to first token

13.76 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 Decode 171.35 · Prefill 761.41 ·
  2. vLLM · FP16 (current)Decode 76.18 · Prefill 254.17 ·
  3. vLLM · IQ3_XS Decode 177.1 · Prefill 886.15 ·

Core figures

Decode (avg)
76.18 tok/s
Prefill (avg)
254.17 tok/s
TTFT (avg)
1.99 s
VRAM (avg)
13.76 GB
MTP acceptance rate
—
TTFB
0.61 GB
Power draw
348.7 W

Configuration

Member-level fields are taken from the most recent run

Model
Llama-3.2-3B-Instruct
Quantization
FP16
Framework
vLLM
Version
v0.47
Context length
2048 tokens
Batch size
8
GPU layers
999
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)
13.76 GB
OS
Windows 11 24H2
Driver
570.86
CUDA
12.6
Power draw
348.7 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

    v0.47 · Windows 11 24H2 · CUDA 12.6 · 2048 ctx

    76.18 tok/s

    Decode

    254.17 tok/s

    Prefill

    1.99 s

    TTFT

    13.76 GB

    VRAM

    —

    MTP

    348.7 W

    Power draw

    L2 batch seed 2026-10-19: vLLM on hw_id=56 running model_id=70 at FP16