Created bymario-alka.dePowered bygodcore.denoob2claw.detricoma.de
Contributed byMario AlkaGoogle

gemma-4-12B-it

Performance benchmark · measured on 29.07.2026 05:40

Benchmark-IDrun-20260729-060805-568759
Timebench 3 - Kombi (Prefill + Generation)Dense12BRuntime: llama.cppQuantisierung: UD-Q4_K_XL
Generation1.058,85tok/s
Prefill3.505,54tok/s
Time to First Token13.703,50ms
Total duration73,68s
Concurrency10parallel
Ranking in the field
13of 29 systems

Performance benchmark · Primary metric: Generation-Speed (tok/s) · 10× concurrent

This run is better than 57 % of all comparable systems.
Generation 1.058,9 tok/s
+10 % vs Ø 960,8
Prefill 3.505,5 tok/s
-67 % vs Ø 10.710,5
Time to First Token 13.704 ms
-65 % vs Ø 38.678
Distribution in the field1 – 2.491 tok/s
Ø 961 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

How does this benchmark compare on other GPUs?

Same model on different hardware · 10× concurrent · Generation (tok/s)

Hardware

GPU: NVIDIA GeForce RTX 5090 · 32 GB VRAM
CPU: AMD Ryzen 7 5800X3D 8-Core Processor
RAM: 126 GB
Mainboard: ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING

Setup

Runtime: llama.cpp
Quantization: UD-Q4_K_XL
Model: gemma-4-12B-it

Configuration

benchmark-konfiguration — run-20260729-060805-568759
# LLM-Benchmark Konfiguration # Modell : gemma-4-12B-it # Engine : llama.cpp # Run-ID : run-20260729-060805-568759 # GPU : NVIDIA GeForce RTX 5090 # CPU : AMD Ryzen 7 5800X3D 8-Core Processor # RAM : 126 GB bench@llm-benchmark:~$ /root/llama.cpp/build/bin/llama-server \ -m /root/.cache/huggingface/hub/models--unsloth--gemma-4-12B-it-qat-GGUF/snapshots/980b060c40a8539ac159e0501a3e0f66a6365af3/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf \ --alias gemma-4-12B-it \ --host 0.0.0.0 \ --port 8000 \ -ngl 999 \ -c 16384 \ -np 4 \ --jinja
Engine?Die Inferenz-Software, die das Modell ausliefert (z.B. vLLM oder llama.cpp). Sie bestimmt Geschwindigkeit, unterstuetzte Modellformate und welche Parameter ueberhaupt verfuegbar sind.llamacpp
Modellalias?Der Name, unter dem das Modell ueber die API angesprochen wird. Genau dieser Wert muss im Request-Feld 'model' stehen.gemma-4-12B-it
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.16384
Modellpfad?Pfad zur GGUF-Modelldatei, die geladen und ausgeliefert wird./root/.cache/huggingface/hub/models--unsloth--gemma-4-12B-it-qat-GGUF/snapshots/980b060c40a8539ac159e0501a3e0f66a6365af3/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf
GPU-Layer?Anzahl der auf die GPU ausgelagerten Modell-Layer. Hoeher = mehr VRAM und schneller; der Rest laeuft auf der CPU. 999 = alles auf GPU.999
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.16384
np4

All benchmarks of this model To leaderboard

Anzeige
Model comparison

gemma-4-12B-it on various hardware

All published performance runs of this model – each bubble a variant: position = prefill (X) × generation (Y), bubble size = number of runs. Closer to the top right = faster. ★ Marked gold = this benchmark.

GPUby graphics card

1.3051.0928806674551.9913.0594.1275.196Prefill (tok/s)Generation (tok/s)NVIDIA RTX PRO 6000 Blackwell Workstation Edition - 1.104,3 tok/s Generation, 4.570 tok/s Prefill, TTFT 5.144 ms (3 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA GeForce RTX 5070 Ti - 677,5 tok/s Generation, 3.217 tok/s Prefill, TTFT 5.812 ms (6 Laufe)NVIDIA GeForce RTX 50...NVIDIA GeForce RTX 3090 Ti - 654,9 tok/s Generation, 2.617 tok/s Prefill, TTFT 8.920 ms (3 Laufe)NVIDIA GeForce RTX 30...NVIDIA GeForce RTX 5090 - 1.058,9 tok/s Generation, 3.140 tok/s Prefill, TTFT 6.383 ms (3 Laufe) | DIESER LAUF★ NVIDIA GeForce RTX 50...
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.104,3 tok/s★ NVIDIA GeForce RTX 5090 1.058,9 tok/s this runNVIDIA GeForce RTX 5070 Ti 677,5 tok/sNVIDIA GeForce RTX 3090 Ti 654,9 tok/s

CPUby processor

1.3051.0928806674551.9913.0594.1275.196Prefill (tok/s)Generation (tok/s)AMD Ryzen 9 9950X 16-Core Processor - 1.104,3 tok/s Generation, 4.570 tok/s Prefill, TTFT 5.144 ms (3 Laufe)AMD Ryzen 9 9950X 16-...AMD Ryzen Threadripper PRO 5975WX 32-Cores - 677,5 tok/s Generation, 3.217 tok/s Prefill, TTFT 5.812 ms (6 Laufe)AMD Ryzen Threadrippe...AMD Ryzen 9 8945HX with Radeon Graphics - 654,9 tok/s Generation, 2.617 tok/s Prefill, TTFT 8.920 ms (3 Laufe)AMD Ryzen 9 8945HX wi...AMD Ryzen 7 5800X3D 8-Core Processor - 1.058,9 tok/s Generation, 3.140 tok/s Prefill, TTFT 6.383 ms (3 Laufe) | DIESER LAUF★ AMD Ryzen 7 5800X3D 8...
AMD Ryzen 9 9950X 16-Core Processor 1.104,3 tok/s★ AMD Ryzen 7 5800X3D 8-Core Processor 1.058,9 tok/s this runAMD Ryzen Threadripper PRO 5975WX 32-Cores 677,5 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 654,9 tok/s

MBby mainboard

1.3051.0928806674551.9913.0594.1275.196Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI - 1.104,3 tok/s Generation, 4.570 tok/s Prefill, TTFT 5.144 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 677,5 tok/s Generation, 3.217 tok/s Prefill, TTFT 5.812 ms (6 Laufe)ASUSTeK COMPUTER INC....Meigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) - 654,9 tok/s Generation, 2.617 tok/s Prefill, TTFT 8.920 ms (3 Laufe)Meigao Innovation Tec...ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING - 1.058,9 tok/s Generation, 3.140 tok/s Prefill, TTFT 6.383 ms (3 Laufe) | DIESER LAUF★ ASUSTeK COMPUTER INC....
ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 1.104,3 tok/s★ ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 1.058,9 tok/s this runASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 677,5 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 654,9 tok/s

ENGby engine

1.3381.0677955242522.9663.3103.6543.997Prefill (tok/s)Generation (tok/s)vLLM - 486,2 tok/s Generation, 3.698 tok/s Prefill, TTFT 3.081 ms (3 Laufe)vLLMllama.cpp - 1.104,3 tok/s Generation, 3.266 tok/s Prefill, TTFT 7.247 ms (12 Laufe) | DIESER LAUF★ llama.cpp
★ llama.cpp 1.104,3 tok/s this runvLLM 486,2 tok/s

DRVby driver

1.2151.1601.1041.0499943.1513.2853.4193.553Prefill (tok/s)Generation (tok/s)unbekannt - 1.104,3 tok/s Generation, 3.352 tok/s Prefill, TTFT 6.414 ms (15 Laufe)unbekannt
unbekannt 1.104,3 tok/s
💰 Economics

Economics of this run

Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (10× concurrent). Methodology →

⚙️ ConfigurationAll metrics and charts below follow these settings – based on a 24-month runtime.Save to URLReset
⚡ Electricity price EUR/kWh
⚙️ System utilization 100 %
🖥️ Acquisition EUR
🔌 Idle 72 W
⚡ TDP 622 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)622 W estimated (TDP)GPU 575 + CPU 35 + Board 12 W full load
Avg cost / hourEUR 0.19
Electricity / 1M tokensEUR 0.049
Token / kWh6.13M
Acquisition (system)EUR 4,986 full priceGPU EUR 3,300 · CPU EUR 349 · Board EUR 149 · RAM EUR 1,008 · PSU EUR 180
Electricity (2 years)
TCO (2 years)EUR 8,253
Output tokens (2 years)66.78B
☁️ External LLM (API) – comparison
External LLM cost (2 years)
Savings vs. external (2 years)

All values above and the charts below take the configured system utilization into account: at X% the system generates only X% of the time, the rest it idles (72 W). Cost per hour drops (more idle), cost per token rises.

Cost over 2 years – electricity only

Cost over 2 years – incl. acquisition (TCO)

Speed vs. tokens per euro

Euro per 1M tokens

Comparison vs. API – economics per benchmark

gemma-4-12B-itNVIDIA GeForce RTX 5090gemma-4-12B-itNVIDIA RTX PRO 6000 Blackwell Workstation Editiongemma-4-12B-itNVIDIA GeForce RTX 5070 Tigemma-4-12B-itNVIDIA GeForce RTX 3090 Ti
Electricity cost (24 mo.)
Acquisition cost
Total cost (TCO)
Generated tokens (24 mo.)
Token price via API
Break-even point (days)
Result (savings / extra cost)

Comparison with up to 3 next-best runs of this model at the same concurrency (at least one on different hardware). Power = GPU TDP + CPU (idle + 15 %) + board (estimated), acquisition = full system (GPU + CPU + board + RAM + PSU), prices = stored market prices.

Contributed by

Mario Alka Administrator

@marioalka

Ich bin Unternehmer, Softwareentwickler und KI-Enthusiast. Seit vielen Jahren entwickle ich Unternehmenssoftware und beschäftige mich inzwischen fast täglich mit lokalen LLMs, KI-Agenten und leistungsfähiger KI-Hardware.

Mit LLM-Benchmark.de möchte ich eine Plattform schaffen, auf der Modelle, GPUs und Agenten objektiv und reproduzierbar miteinander verglichen werden.