Contributed byMario AlkaQwen (Alibaba) Qwen3-Coder-30B-A3B-Instruct
Performance benchmark · measured on 18.08.2026 14:25
Benchmark-ID
run-20260818-144651-c4773bTimebench 3 - Kombi (Prefill + Generation)MoE30BRuntime: llama.cppQuantisierung: Q4_K_M
Generation276,95tok/s
Prefill7.786,17tok/s
Time to First Token4.142,50ms
Total duration83,22s
Concurrency10parallel
Ranking in the field
663of 1427 systems
Performance benchmark · Primary metric: Generation-Speed (tok/s) · 10× concurrent
This run is better than 54 % of all comparable systems.
Generation
277,0 tok/s
-35 % vs Ø 425,3
Prefill
7.786,2 tok/s
+17 % vs Ø 6.659,3
Time to First Token
4.143 ms
-91 % vs Ø 45.848
Distribution in the field0 – 2.491 tok/s
Wie schlägt sich dieser Benchmark mit anderen Modellen?
gemma-4-E2B-itNVIDIA GeForce RTX 5090 · run-20260728-184455-e28775
2.491,2 tok/s
gemma-4-E2B-itNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181606-b37ed9
2.414,4 tok/s
Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 Ti · run-20261021-093815-dce558
2.263,7 tok/s
Nemotron-3-Nano-4BNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260727-181607-4cca42
2.182,7 tok/s
gemma-4-E2B-it3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-184456-d87c73
2.143,6 tok/s
gpt-oss-20bNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260730-035052-9c966e
1.995,4 tok/s
gpt-oss-20bNVIDIA RTX PRO 6000 Blackwell Workstation Edition · run-20260728-140954-8007c4
1.993,5 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-0aed86
1.937,7 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-8db0fa
1.911,6 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260727-181606-e0eafd
1.897,0 tok/s
gpt-oss-20bNVIDIA GeForce RTX 5090 · run-20260729-032121-7c961d
1.890,7 tok/s
Nemotron-3-Nano-4BNVIDIA GeForce RTX 5090 · run-20260728-194135-2ed9dd
1.878,7 tok/s
gpt-oss-20b3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-034052-26063b
1.835,7 tok/s
Nemotron-3-Nano-4B3× 3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-a77046
1.830,3 tok/s
Qwen3-Coder-30B-A3B-Instruct this run2× 2x AMD Radeon PRO W7900 Dual Slot · run-20260818-144651-c4773b
277,0 tok/s
How does this benchmark compare on other GPUs?
Same model on different hardware · 10× concurrent · Generation (tok/s)
NVIDIA GeForce RTX 5090run-20260729-105335-622167
1.716,6 tok/s
NVIDIA RTX PRO 6000 Blackwell Workstation Editionrun-20260729-032124-ef2bb1
1.690,0 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-105343-8e8f9f
1.545,3 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-105342-c6ca9d
1.531,9 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-105342-11e651
1.455,4 tok/s
NVIDIA RTX PRO 6000 Blackwell Workstation Editionrun-20260724-004606-ebf474
1.070,6 tok/s
NVIDIA GeForce RTX 5090run-20260723-045450-366dae
966,2 tok/s
NVIDIA GeForce RTX 3090 Tirun-20260729-060804-8652f4
928,5 tok/s
2× NVIDIA RTX A6000run-20260829-120719-f95c4a
789,2 tok/s
2× NVIDIA RTX A6000run-20260827-083602-a5482a
720,9 tok/s
NVIDIA GeForce RTX 3090 Tirun-20260724-004603-59220f
670,7 tok/s
2× NVIDIA RTX A6000run-20260827-203813-2d45d6
593,7 tok/s
2× NVIDIA RTX A6000run-20260829-181042-fa807e
582,7 tok/s
2× NVIDIA RTX A6000run-20260830-033127-15b23f
562,8 tok/s
Hardware
GPU: 2x AMD Radeon PRO W7900 Dual Slot · 48 GB VRAM
CPU: AMD Ryzen Threadripper PRO 5975WX 32-Cores
RAM: 247 GB
Mainboard: ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI
Setup
Runtime: llama.cpp
Quantization: Q4_K_M
Model: Qwen3-Coder-30B-A3B-Instruct
Configuration
# LLM-Benchmark Konfiguration
# Modell : Qwen3-Coder-30B-A3B-Instruct
# Engine : llama.cpp
# Run-ID : run-20260818-144651-c4773b
# GPU : 2x AMD Radeon PRO W7900 Dual Slot
# CPU : AMD Ryzen Threadripper PRO 5975WX 32-Cores
# RAM : 247 GB
bench@llm-benchmark:~$ /home/godcore/llama.cpp/build_hip/bin/llama-server \
-m /home/godcore/bench_scratch/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf \
--alias Qwen3-Coder-30B-A3B-Instruct \
-ngl 999 \
-fa on \
-sm layer \
--tensor-split 1,1 \
-c 49152 \
-np 12 \
--jinja \
--host 0.0.0.0 \
--port 8000
| 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. | Qwen3-Coder-30B-A3B-Instruct |
| Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann. | 49152 |
| Modellpfad?Pfad zur GGUF-Modelldatei, die geladen und ausgeliefert wird. | /home/godcore/bench_scratch/Qwen3-Coder-30B-A3B-Instruct-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 |
| fa | on |
| Split-Mode?Verteilung ueber mehrere GPUs: none (nur eine GPU), layer (Layer aufteilen) oder row (Tensoren zeilenweise). | layer |
| Tensor-Split?Verhaeltnis, in dem die Modell-Layer auf mehrere GPUs verteilt werden, z.B. 3,1. | 1,1 |
| Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells. | 49152 |
| np | 12 |
| execution_type | local |
Anzeige
Model comparison
Qwen3-Coder-30B-A3B-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
NVIDIA GeForce RTX 5090 1.716,6 tok/sNVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.690,0 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1.545,3 tok/sNVIDIA GeForce RTX 3090 Ti 928,5 tok/sNVIDIA RTX A6000 789,2 tok/sNVIDIA GeForce RTX 4090 400,4 tok/sAMD Radeon RX 7900 XTX 324,9 tok/sAMD Radeon PRO W7800 48GB 289,5 tok/s★ AMD Radeon PRO W7900 Dual Slot 277,0 tok/s this runAMD Radeon AI PRO R9700 250,2 tok/sNVIDIA GB10 (DGX Spark) 182,9 tok/sNVIDIA GeForce RTX 5070 Ti 157,0 tok/sCPU-only 6,7 tok/s
CPUby processor
AMD Ryzen 7 5800X3D 8-Core Processor 1.716,6 tok/sAMD Ryzen 9 9950X 16-Core Processor 1.690,0 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 1.545,3 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 928,5 tok/sAMD Ryzen Threadripper PRO 7955WX 16-Cores 789,2 tok/sAMD Ryzen 3 3100 4-Core Processor 400,4 tok/sAMD Ryzen Threadripper PRO 3955WX 16-Cores 324,9 tok/s★ AMD Ryzen Threadripper PRO 5975WX 32-Cores 289,5 tok/s this runIntel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz 234,1 tok/sNVIDIA Grace 182,9 tok/sIntel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz 6,7 tok/s
MBby mainboard
ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 1.716,6 tok/sASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 1.690,0 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 1.545,3 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 928,5 tok/sMicro-Star International Co., Ltd. B550M PRO-VDH WIFI (MS-7C95) 400,4 tok/s★ ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 324,9 tok/s this runASUSTeK COMPUTER INC. ROG STRIX X570-F GAMING 250,2 tok/sDell Inc. PowerEdge R820 234,1 tok/sASUSTeK COMPUTER INC. GX10 182,9 tok/s
DRVby driver
unbekannt 1.716,6 tok/sNVIDIA 590.48.01 / CUDA 13.1 182,9 tok/sAMD 7.0.0-27-generic 88,2 tok/s
💰 Economics
Economics of this run
Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (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
10 W
⚡ TDP
600 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)600 W estimated (TDP)GPU 590 + Board 10 W full load
Avg cost / hourEUR 0.18
Electricity / 1M tokensEUR 0.18
Token / kWh1.66M
Acquisition (system)EUR 11,858 partial priceGPU EUR 7,822 · CPU EUR 1,880 · RAM EUR 1,976 · PSU EUR 180
Electricity (2 years)–
TCO (2 years)EUR 15,012
Output tokens (2 years)17.47B
☁️ 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 (10 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
| Qwen3-Coder-30B-A3B-Instruct2x AMD Radeon PRO W7900 Dual Slot | Qwen3-Coder-30B-A3B-InstructNVIDIA GeForce RTX 5090 | Qwen3-Coder-30B-A3B-InstructNVIDIA RTX PRO 6000 Blackwell Workstation Edition | Qwen3-Coder-30B-A3B-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) | – | – | – | – |
2× AMD Radeon PRO W7900 Dual Slot · this run 277 tok/sNVIDIA GeForce RTX 5090 1,717 tok/sNVIDIA RTX PRO 6000 Blackwell Workstation Edition 1,690 tok/s3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1,545 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.