Llama-3.1-8B-Instruct · Q4_K_M
This page aggregates 2 real-world runs of Llama-3.1-8B-Instruct (Q4_K_M) on NVIDIA RTX 3090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
116.04 tok/s
Decode
Decode speed
765.01 tok/s
Prefill
Prefill speed
0.74 s
TTFT
Time to first token
10.64 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2026-10-18
Last verified
Performance
- vLLM · Q8_0 Decode 64.44 · Prefill 290.44 ·
- Ollama · IQ3_XS Decode 101.05 · Prefill 370.05 ·
- llama.cpp · IQ3_XS Decode 111.47 · Prefill 506.1 ·
- llama.cpp · Q4_K_M (current)Decode 116.04 · Prefill 765.01 ·
- Ollama · IQ2_XS Decode 106.45 · Prefill 486.76 ·
- llama.cpp · Q8_0 Decode 84.9 · Prefill 636.8 ·
- llama.cpp · Q5_K_M Decode 123.7 · Prefill 927.8 ·
Core figures
- Decode (avg)
- 116.04 tok/s
- Prefill (avg)
- 765.01 tok/s
- TTFT (avg)
- 0.74 s
- VRAM (avg)
- 10.64 GB
- MTP acceptance rate
- —
- TTFB
- 0.66 GB
- Power draw
- 256.2 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Llama-3.1-8B-Instruct
- Quantization
- Q4_K_M
- Framework
- llama.cpp
- Version
- v0.44
- Context length
- 2048 tokens
- Batch size
- 512
- 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 RTX 3090
- Nominal VRAM
- 24 GB
- Measured VRAM (avg)
- 10.64 GB
- OS
- Ubuntu 22.04
- Driver
- 565.57
- CUDA
- 12.6
- Power draw
- 256.2 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.44 · Ubuntu 22.04 · CUDA 12.6 · 2048 ctx
83.07 tok/s
Decode
412.52 tok/s
Prefill
1.4 s
TTFT
15.88 GB
VRAM
—
MTP
256.2 W
Power draw
L2 batch seed 2026-10-18: llama.cpp on hw_id=2 running model_id=68 at Q4_K_M
b4200 · Linux · CUDA 12.6 · 4096 ctx
149 tok/s
Decode
1,117.5 tok/s
Prefill
0.09 s
TTFT
5.4 GB
VRAM
—
MTP
— W
Power draw
L0 理论估算(add-1000-records-batch3)