Llama-3.1-8B-Instruct · Q5_K_M
This page aggregates 2 real-world runs of Llama-3.1-8B-Instruct (Q5_K_M) on NVIDIA RTX 4090 with llama.cpp, contributed by 1 independent source platforms; metrics are averages of published measurements.
128.55 tok/s
Decode
Decode speed
827.45 tok/s
Prefill
Prefill speed
0.96 s
TTFT
Time to first token
12.06 GB
VRAM
VRAM usage
2
Measured runs
1
Independent sources
Other
Source platforms
2026-10-18
Last verified
Performance
- vLLM · IQ2_XS Decode 205.12 · Prefill 1126.08 ·
- Strata · Q4_K_M Decode 191.74 · Prefill 642.2 ·
- Strata · Q6_K Decode 151.86 · Prefill 574.29 ·
- llama.cpp · Q5_K_M (current)Decode 128.55 · Prefill 827.45 ·
- vLLM · Q4_K_M Decode 142.74 · Prefill 510.45 ·
- llama.cpp · Q8_0 Decode 91.4 · Prefill 685.5 ·
- llama.cpp · Q4_K_M Decode 160.5 · Prefill 1203.8 ·
Core figures
- Decode (avg)
- 128.55 tok/s
- Prefill (avg)
- 827.45 tok/s
- TTFT (avg)
- 0.96 s
- VRAM (avg)
- 12.06 GB
- MTP acceptance rate
- —
- TTFB
- 0.99 GB
- Power draw
- 331.6 W
Configuration
Member-level fields are taken from the most recent run
- Model
- Llama-3.1-8B-Instruct
- Quantization
- Q5_K_M
- Framework
- llama.cpp
- Version
- v0.49
- Context length
- 32768 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 RTX 4090
- Nominal VRAM
- 24 GB
- Measured VRAM (avg)
- 12.06 GB
- OS
- Ubuntu 24.04
- Driver
- 560.94
- CUDA
- 13.0
- Power draw
- 331.6 W
Sources & evidence
2 measured records in total, each traceable to its original source
v0.49 · Ubuntu 24.04 · CUDA 13.0 · 32768 ctx
123.8 tok/s
Decode
655.09 tok/s
Prefill
1.82 s
TTFT
17.72 GB
VRAM
—
MTP
331.6 W
Power draw
L2 batch seed 2026-10-18: llama.cpp on hw_id=1 running model_id=68 at Q5_K_M
b4200 · Linux · CUDA 12.6 · 4096 ctx
133.3 tok/s
Decode
999.8 tok/s
Prefill
0.1 s
TTFT
6.4 GB
VRAM
—
MTP
— W
Power draw
L0 理论估算(add-1000-records-batch3)