Created bymario-alka.dePowered bygodcore.denoob2claw.detricoma.de
Contributed byMario AlkaQwen (Alibaba)

Qwen2.5-72B-Instruct

Performance benchmark · measured on 20.07.2026 14:56

Benchmark-IDrun-20260722-165905-1dde01
Performancetest Small 1.0Dense72BRuntime: llama.cppQuantisierung: Q4_K_M
Generation2,28tok/s
Prefill80,82tok/s
Time to First Token569,00ms
Total duration51,82s
Concurrency1parallel
Ranking in the field
28of 33 systems

Performance benchmark · Primary metric: Generation-Speed (tok/s) · 1× concurrent

This run is better than 16 % of all comparable systems.
Generation 2,3 tok/s
-90 % vs Ø 21,9
Prefill 80,8 tok/s
-87 % vs Ø 605,3
Time to First Token 569 ms
-99 % vs Ø 44.665
Distribution in the field1 – 81 tok/s
Ø 22 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

How does this benchmark compare on other GPUs?

Same model on different hardware · 1× concurrent · Generation (tok/s)

Hardware

GPU: 2x NVIDIA GeForce RTX 2060 · 6 GB VRAM
CPU: 4x Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz
RAM: 504 GB
Mainboard: Dell Inc. PowerEdge R820

Setup

Runtime: llama.cpp
Quantization: Q4_K_M
Model: Qwen2.5-72B-Instruct

Configuration

benchmark-konfiguration — run-20260722-165905-1dde01
# LLM-Benchmark Konfiguration # Modell : Qwen2.5-72B-Instruct # Engine : llama.cpp # Run-ID : run-20260722-165905-1dde01 # GPU : 2x NVIDIA GeForce RTX 2060 # CPU : 4x Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz # RAM : 504 GB bench@llm-benchmark:~$ /opt/llama.cpp/build/bin/llama-server \ -m /opt/models/Qwen2.5-72B-Instruct-Q4_K_M.gguf \ -a Qwen2.5-72B-Instruct \ --host 0.0.0.0 \ --port 8080 \ --numa distribute \ -t 64 \ -tb 64 \ -c 8192 \ --parallel 4
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.Qwen2.5-72B-Instruct
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.8192
Modellpfad?Pfad zur GGUF-Modelldatei, die geladen und ausgeliefert wird./opt/models/Qwen2.5-72B-Instruct-Q4_K_M.gguf
Alias?Anzeigename des Modells nach aussen (served model name), unabhaengig vom Dateinamen.Qwen2.5-72B-Instruct
NUMA?NUMA-Optimierung fuer Multi-Socket-CPUs: distribute/isolate/numactl. Verbessert die Speicherlokalitaet.distribute
Threads?Anzahl CPU-Threads fuer die Token-Generierung (Decode).64
Batch-Threads?Anzahl CPU-Threads fuer Prompt-Verarbeitung und Batch (Prefill).64
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.8192
Parallel?Anzahl paralleler Slots/Sequenzen, die der Server gleichzeitig bedient. Der Kontext wird auf die Slots aufgeteilt.4

All benchmarks of this model To leaderboard

Anzeige
Model comparison

Qwen2.5-72B-Instruct 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

31123315577,60,009391.8782.818Prefill (tok/s)Generation (tok/s)NVIDIA RTX PRO 6000 Blackwell Workstation Edition - 239,2 tok/s Generation, 2.275 tok/s Prefill, TTFT 19.196 ms (3 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition - 203,8 tok/s Generation, 2.269 tok/s Prefill, TTFT 23.754 ms (9 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA GeForce RTX 5070 Ti - 5,8 tok/s Generation, 159 tok/s Prefill, TTFT 42.383 ms (3 Laufe)NVIDIA GeForce RTX 50...NVIDIA GeForce RTX 3090 Ti - 4,9 tok/s Generation, 209 tok/s Prefill, TTFT 31.676 ms (3 Laufe)NVIDIA GeForce RTX 30...NVIDIA GeForce RTX 5090 - 3,0 tok/s Generation, 147 tok/s Prefill, TTFT 42.799 ms (3 Laufe)NVIDIA GeForce RTX 50...CPU-only - 2,3 tok/s Generation, 81 tok/s Prefill, TTFT 569 ms (1 Lauf)CPU-onlyAMD Radeon AI PRO R9700 - 1,8 tok/s Generation, 19 tok/s Prefill, TTFT 300.177 ms (9 Laufe)AMD Radeon AI PRO R97...NVIDIA GeForce RTX 2060 - 2,3 tok/s Generation, 81 tok/s Prefill, TTFT 569 ms (1 Lauf) | DIESER LAUF★ NVIDIA GeForce RTX 20...
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 239,2 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 203,8 tok/sNVIDIA GeForce RTX 5070 Ti 5,8 tok/sNVIDIA GeForce RTX 3090 Ti 4,9 tok/sNVIDIA GeForce RTX 5090 3,0 tok/sCPU-only 2,3 tok/s★ NVIDIA GeForce RTX 2060 2,3 tok/s this runAMD Radeon AI PRO R9700 1,8 tok/s

CPUby processor

31123315577,60,009391.8782.818Prefill (tok/s)Generation (tok/s)AMD Ryzen 9 9950X 16-Core Processor - 239,2 tok/s Generation, 2.275 tok/s Prefill, TTFT 19.196 ms (3 Laufe)AMD Ryzen 9 9950X 16-...AMD Ryzen Threadripper PRO 9965WX 24-Cores - 203,8 tok/s Generation, 2.269 tok/s Prefill, TTFT 23.754 ms (9 Laufe)AMD Ryzen Threadrippe...AMD Ryzen Threadripper PRO 5975WX 32-Cores - 5,8 tok/s Generation, 159 tok/s Prefill, TTFT 42.383 ms (3 Laufe)AMD Ryzen Threadrippe...AMD Ryzen 9 8945HX with Radeon Graphics - 4,9 tok/s Generation, 209 tok/s Prefill, TTFT 31.676 ms (3 Laufe)AMD Ryzen 9 8945HX wi...AMD Ryzen 7 5800X3D 8-Core Processor - 3,0 tok/s Generation, 147 tok/s Prefill, TTFT 42.799 ms (3 Laufe)AMD Ryzen 7 5800X3D 8...Intel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz - 2,3 tok/s Generation, 81 tok/s Prefill, TTFT 569 ms (1 Lauf)Intel(R) Xeon(R) CPU ...AMD Ryzen Threadripper PRO 7955WX 16-Cores - 1,8 tok/s Generation, 19 tok/s Prefill, TTFT 300.177 ms (9 Laufe)AMD Ryzen Threadrippe...Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz - 2,3 tok/s Generation, 81 tok/s Prefill, TTFT 569 ms (1 Lauf) | DIESER LAUF★ Intel(R) Xeon(R) CPU ...
AMD Ryzen 9 9950X 16-Core Processor 239,2 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 203,8 tok/sAMD Ryzen Threadripper PRO 5975WX 32-Cores 5,8 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 4,9 tok/sAMD Ryzen 7 5800X3D 8-Core Processor 3,0 tok/sIntel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz 2,3 tok/s★ Intel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz 2,3 tok/s this runAMD Ryzen Threadripper PRO 7955WX 16-Cores 1,8 tok/s

MBby mainboard

31023315577,60,009361.8712.807Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI - 239,2 tok/s Generation, 2.275 tok/s Prefill, TTFT 19.196 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE - 203,8 tok/s Generation, 1.144 tok/s Prefill, TTFT 161.966 ms (18 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 5,8 tok/s Generation, 159 tok/s Prefill, TTFT 42.383 ms (3 Laufe)ASUSTeK COMPUTER INC....Meigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) - 4,9 tok/s Generation, 209 tok/s Prefill, TTFT 31.676 ms (3 Laufe)Meigao Innovation Tec...ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING - 3,0 tok/s Generation, 147 tok/s Prefill, TTFT 42.799 ms (3 Laufe)ASUSTeK COMPUTER INC....Dell Inc. PowerEdge R820 - 2,3 tok/s Generation, 81 tok/s Prefill, TTFT 569 ms (2 Laufe) | DIESER LAUF★ Dell Inc. PowerEdge R...
ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 239,2 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 203,8 tok/sASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 5,8 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 4,9 tok/sASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 3,0 tok/s★ Dell Inc. PowerEdge R820 2,3 tok/s this run

ENGby engine

31023315577,60,003827651.147Prefill (tok/s)Generation (tok/s)unbekannt - 2,3 tok/s Generation, 81 tok/s Prefill, TTFT 569 ms (1 Lauf)unbekanntllama.cpp - 239,2 tok/s Generation, 937 tok/s Prefill, TTFT 107.229 ms (31 Laufe) | DIESER LAUF★ llama.cpp
★ llama.cpp 239,2 tok/s this rununbekannt 2,3 tok/s

DRVby driver

263251239227215856892928965Prefill (tok/s)Generation (tok/s)unbekannt - 239,2 tok/s Generation, 910 tok/s Prefill, TTFT 103.896 ms (32 Laufe)unbekannt
unbekannt 239,2 tok/s
💰 Economics

Economics of this run

Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (1× 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 55 W
⚡ TDP 359 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)359 W estimated (TDP)GPU 320 + CPU 29 + Board 10 W full load
Avg cost / hourEUR 0.11
Electricity / 1M tokensEUR 13.11
Token / kWh22.88K
Acquisition (system)EUR 2,445 partial priceGPU EUR 220 · CPU EUR 59 · RAM EUR 2,016 · PSU EUR 150
Electricity (2 years)
TCO (2 years)EUR 4,331
Output tokens (2 years)143.80M
☁️ 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 (55 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

Qwen2.5-72B-Instruct2x NVIDIA GeForce RTX 2060Qwen2.5-72B-InstructNVIDIA RTX PRO 6000 Blackwell Workstation EditionQwen2.5-72B-Instruct3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionQwen2.5-72B-Instruct3x 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)

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.