Card × model benchmark

Phi-4 on RTX 4070 Ti Q4_K_M

Comfortable

On an RTX 4070 Ti (12 GB), Phi-4 in Q4_K_M generates 50,8 tokens/s and reads the prompt at 3 442 tokens/s, using 9,2 Go of video memory.

Median of 1 run · latest on 25/09/26

50,8 tok/s

Generation

3 442 tok/s

Prompt processing

9,2 Go

VRAM used (peak)

171 W

Average power

0,30 tok/s/W

Efficiency

0,23 €/M tok

Electricity / M tokens · 0,25 €/kWh

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

RTX 4070 Ti page All cards on Phi-4 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 12 GB?	Estimated speed
4 096	0,8 Go	9,9 Go	yes	≈ 47,3 tok/s
8 192	1,6 Go	10,6 Go	yes	≈ 43,8 tok/s
16 384	3,1 Go	12,2 Go	no	—

KV cache per token = 2 × 40 layers × 10 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”

Phi-4 on other cards

Card	VRAM	Generation	Difference	Price
RTX 5070 Ti	16 Go	84,9 tok/s	+67,2 %	—
RTX 4070 Ti	12 Go	50,8 tok/s	this card	—

Other models measured on the RTX 4070 Ti

Llama 3.2 3B Instruct Q4_K_M 189 tok/s Qwen2.5 7B Instruct Q4_K_M 98,8 tok/s Qwen2.5 Coder 7B Q4_K_M 96,7 tok/s Meta Llama 3.1 8B Instruct Q4_K_M 91,9 tok/s Qwen2.5 14B Instruct Q4_K_M 50,5 tok/s DeepSeek R1 Distill Qwen 14B Q4_K_M 51,1 tok/s Qwen2.5 Coder 14B Q4_K_M 50,3 tok/s

All measurements

Date	Source	Type	Generation	Prompt	VRAM	Max temp.	Power	llama.cpp
25/09/26	SephiGame	context 16 384ASUS ROG Strix	4,5	92,8	11,6 Go	51 °C	76 W	llama.cpp 96278e39f
25/09/26	SephiGame	context 8 192ASUS ROG Strix	43,0	2 729	10,8 Go	67 °C	192 W	llama.cpp 96278e39f
25/09/26	SephiGame	standardASUS ROG Strix	50,8	3 442	9,2 Go	66 °C	171 W	llama.cpp 96278e39f

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