Contributed byMario AlkaGoogle gemma-4-E2B-it
Performance benchmark · measured on 29.07.2026 21:49
Benchmark-ID
run-20260729-224611-fd132dTimebench 3 - Kombi (Prefill + Generation)Dense5BRuntime: llama.cppQuantisierung: Q4_K_M
ℹ
For context: Diese Plattform nutzt Unified Memory – die "VRAM" ist gemeinsamer System-RAM (APU/Superchip); das Modell teilt sich den Speicher mit dem System.
Generation461,54tok/s
Prefill5.838,37tok/s
Time to First Token3.794,50ms
Total duration44,75s
Concurrency5parallel
Ranking in the field
2of 6 systems
Performance benchmark · Primary metric: Generation-Speed (tok/s) · 5× concurrent
This run is better than 80 % of all comparable systems.
Generation
461,5 tok/s
+71 % vs Ø 269,4
Prefill
5.838,4 tok/s
+16 % vs Ø 5.013,5
Time to First Token
3.795 ms
-38 % vs Ø 6.118
Distribution in the field78 – 479 tok/s
Wie schlägt sich dieser Benchmark mit anderen Modellen?
gemma-4-E2B-it same modelNVIDIA GB10 · run-20260729-224610-d54f79
478,6 tok/s
gemma-4-E2B-it this runNVIDIA GB10 · run-20260729-224611-fd132d
461,5 tok/s
gemma-4-E4B-itNVIDIA GB10 · run-20260729-224611-1ec61a
246,7 tok/s
gemma-4-E4B-itNVIDIA GB10 · run-20260729-224611-f3c242
245,6 tok/s
MiniMax-M2.7NVIDIA GB10 · run-20260723-055735-105350
106,1 tok/s
MiniMax-M2.5NVIDIA GB10 · run-20260726-091400-7d7441
77,7 tok/s
How does this benchmark compare on other GPUs?
Same model on different hardware · 5× concurrent · Generation (tok/s)
NVIDIA RTX PRO 6000 Blackwell Workstation Editionrun-20260727-181606-0adf6d
1.349,3 tok/s
NVIDIA GeForce RTX 5090run-20260728-184455-5b4937
1.301,9 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260728-184456-4111a9
1.195,7 tok/s
NVIDIA GeForce RTX 5070 Tirun-20260727-032758-a7d32d
973,5 tok/s
NVIDIA GeForce RTX 5070 Tirun-20260727-090152-b472f0
971,4 tok/s
NVIDIA GeForce RTX 5070 Tirun-20260728-140953-67115b
969,6 tok/s
NVIDIA GeForce RTX 3090 Tirun-20260727-142111-0dd411
827,3 tok/s
NVIDIA GB10run-20260729-224610-d54f79
478,6 tok/s
NVIDIA GB10 this runrun-20260729-224611-fd132d
461,5 tok/s
2× AMD Radeon Graphicsrun-20260729-224611-5817c5
436,3 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260727-155936-966bd8
156,1 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260728-184456-617786
153,6 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260728-184456-0683d8
144,4 tok/s
Hardware
GPU: NVIDIA GB10 · 128 GB VRAM
CPU: NVIDIA Grace
RAM: 120 GB
Mainboard: ASUSTeK COMPUTER INC. GX10
Setup
Runtime: llama.cpp
Quantization: Q4_K_M
Driver: NVIDIA 590.48.01 / CUDA 13.1
Model: gemma-4-E2B-it
Configuration
# LLM-Benchmark Konfiguration
# Modell : gemma-4-E2B-it
# Engine : llama.cpp
# Run-ID : run-20260729-224611-fd132d
# GPU : NVIDIA GB10
# CPU : NVIDIA Grace
# RAM : 120 GB
bench@llm-benchmark:~$ /home/godcore/llama.cpp/build/bin/llama-server \
-m /home/godcore/.cache/huggingface/hub/models--unsloth--gemma-4-E2B-it-GGUF/snapshots/0314792d7f1f7e229411f620751375812bb9faf2/gemma-4-E2B-it-Q4_K_M.gguf \
--alias gemma-4-E2B-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-E2B-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. | /home/godcore/.cache/huggingface/hub/models--unsloth--gemma-4-E2B-it-GGUF/snapshots/0314792d7f1f7e229411f620751375812bb9faf2/gemma-4-E2B-it-Q4_K_M.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 |
| np | 4 |
Anzeige
Model comparison
gemma-4-E2B-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
NVIDIA GeForce RTX 5090 2.491,2 tok/sNVIDIA RTX PRO 6000 Blackwell Workstation Edition 2.414,4 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 2.143,6 tok/sNVIDIA GeForce RTX 5070 Ti 1.744,0 tok/sNVIDIA GeForce RTX 3090 Ti 1.492,8 tok/s★ NVIDIA GB10 (DGX Spark) 903,3 tok/s this runAMD Radeon Graphics 826,2 tok/s
CPUby processor
AMD Ryzen 7 5800X3D 8-Core Processor 2.491,2 tok/sAMD Ryzen 9 9950X 16-Core Processor 2.414,4 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 2.143,6 tok/sAMD Ryzen Threadripper PRO 5975WX 32-Cores 1.744,0 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 1.492,8 tok/s★ NVIDIA Grace 903,3 tok/s this runAMD Ryzen 9 7945HX with Radeon Graphics 826,2 tok/s
MBby mainboard
ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 2.491,2 tok/sASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 2.414,4 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 2.143,6 tok/sASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 1.744,0 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 1.492,8 tok/s★ ASUSTeK COMPUTER INC. GX10 903,3 tok/s this runShenzhen Meigao Electronic Equipment Co.,Ltd DRFXI (MotherBoard Series) 826,2 tok/s
DRVby driver
unbekannt 2.491,2 tok/s★ NVIDIA 590.48.01 / CUDA 13.1 903,3 tok/s this run
💰 Economics
Economics of this run
Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (5× 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
15 W
⚡ TDP
140 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)140 W estimated (TDP)GPU 140 W full load
Avg cost / hourEUR 0.042
Electricity / 1M tokensEUR 0.025
Token / kWh11.87M
Acquisition (system)EUR 5,800 partial priceGPU EUR 4,000 · RAM EUR 1,680 · PSU EUR 120
Electricity (2 years)–
TCO (2 years)EUR 6,536
Output tokens (2 years)29.11B
☁️ 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 (15 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-E2B-itNVIDIA GB10 (DGX Spark) | gemma-4-E2B-itNVIDIA RTX PRO 6000 Blackwell Workstation Edition | gemma-4-E2B-itNVIDIA GeForce RTX 5090 | gemma-4-E2B-it3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | |
|---|---|---|---|---|
| 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) | – | – | – | – |
NVIDIA GB10 (DGX Spark) · this run 462 tok/sNVIDIA RTX PRO 6000 Blackwell Workstation Edition 1,349 tok/sNVIDIA GeForce RTX 5090 1,302 tok/s3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1,196 tok/s
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.