Card × model benchmark

Meta Llama 3.1 8B Instruct on RTX 4090 Q4_K_M

Smooth

On an RTX 4090 (24 GB), Meta Llama 3.1 8B Instruct in Q4_K_M generates 132 tokens/s and reads the prompt at 5 900 tokens/s, using 6,2 Go of video memory.

Median of 1 run · latest on 04/08/26

Demo data: this combination has not actually been measured yet.

132 tok/s

Generation

5 900 tok/s

Prompt processing

6,2 Go

VRAM used (peak)

285 W

Average power

0,46 tok/s/W

Efficiency

0,15 €/M tok

Electricity / M tokens · 0,25 €/kWh

Estimate your monthly cost from your electricity price and usage: sign in

RTX 4090 page All cards on Meta Llama 3.1 8B Instruct Feel the speed Rent in the cloud

Where to buyLDLC Materiel.net TopAchat Cybertek Grosbill Rue du Commerce Amazon

What about a longer context?

An estimate, not a measurement: the KV cache grows with context, takes VRAM and slows generation down.

Context	KV cache	Estimated VRAM	Fits in 24 GB?	Estimated speed
4 096	0,5 Go	6,6 Go	yes	≈ 122 tok/s
8 192	1,0 Go	7,1 Go	yes	≈ 111 tok/s
16 384	2,0 Go	8,1 Go	yes	≈ 94,9 tok/s
32 768	4,0 Go	10,1 Go	yes	≈ 73,5 tok/s
65 536	8,0 Go	14,1 Go	yes	≈ 50,6 tok/s
131 072	16,0 Go	22,1 Go	yes	≈ 31,2 tok/s

KV cache per token = 2 × 32 layers × 8 KV heads × 128 × 2 bytes (f16). Estimated VRAM = measured VRAM + added KV cache. Speed ≈ measured speed × weights ÷ (weights + added KV cache) — attention compute is not modelled, the real drop is slightly larger. Fine-tune it in the “Configurator”

Meta Llama 3.1 8B Instruct on other cards

Card	VRAM	Generation	Difference	Price
RTX 5070 Ti	16 Go	151 tok/s	+14,5 %	1 400 €
RTX 4090demo	24 Go	132 tok/s	this card	—
RTX 5080demo	16 Go	124 tok/s	−6,1 %	1 850 €
RTX 4080demo	16 Go	103 tok/s	−22,0 %	780 €
RTX 4070 Ti	12 Go	89,6 tok/s	−32,1 %	927 €
RTX 4070	12 Go	82,3 tok/s	−37,7 %	595 €
RTX 3060demo	12 Go	48,0 tok/s	−63,6 %	260 €

Other models measured on the RTX 4090

Qwen2.5 14B Instruct Q4_K_M 78,0 tok/s

All measurements

Date	Source	Type	Generation	Prompt	VRAM	Max temp.	Power	llama.cpp
04/08/26	demo	standard	132	5 900	6,2 Go	66 °C	285 W	—

Protocol: llama-bench (llama.cpp), prompt 512 + generation 128 tokens, model fully loaded in VRAM. VRAM used = system peak (includes the OS, ~0.5 to 1 GB more than the model alone). Measure my card