ℹ

For context: Teile des Modells wurden per --cpu-offload-gb=32 GB in den System-RAM ausgelagert – RAM wird als erweiterter VRAM genutzt, was den Durchsatz senken kann.

Generation18,75tok/s

Prefill986,99tok/s

Time to First Token26.526,50ms

Total duration1.146,34s

Concurrency10parallel

Ranking in the field

This GPU All GPUs

175of 175 systems

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

This run is better than 0 % of all comparable systems.

Generation 18,8 tok/s

-98 % vs Ø 954,8

Prefill 987,0 tok/s

-87 % vs Ø 7.601,4

Time to First Token 26.527 ms

+33 % vs Ø 19.972

Distribution in the field19 – 2.144 tok/s

Ø 955 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

gemma-4-E2B-it3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-184456-d87c73

2.143,6 tok/s

Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-0aed86

1.937,7 tok/s

Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-8db0fa

1.911,6 tok/s

Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260727-181606-e0eafd

1.897,0 tok/s

gpt-oss-20b3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-034052-26063b

1.835,7 tok/s

Nemotron-3-Nano-4B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260728-194136-a77046

1.830,3 tok/s

gpt-oss-20b3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-032121-c6e7e0

1.819,5 tok/s

gpt-oss-20b3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-032121-2b99aa

1.802,0 tok/s

Ornith-1.0-35B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260731-071556-e9a354

1.799,4 tok/s

gpt-oss-20b3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-032120-1cc62d

1.742,1 tok/s

Nemotron-Cascade-2-30B-A3B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-105340-3fa079

1.655,2 tok/s

Nemotron-3-Nano-30B-A3B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-105341-be087a

1.654,6 tok/s

Nemotron-3-Nano-Omni-30B-A3B-Reasoning3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-105340-2fab14

1.651,0 tok/s

Nemotron-Cascade-2-30B-A3B3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260729-105340-b8a790

1.650,7 tok/s

GLM-4.5-Air this runNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · run-20260924-131556-ded9b8

18,8 tok/s

How does this benchmark compare on other GPUs?

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

NVIDIA RTX PRO 6000 Blackwell Workstation Editionrun-20260728-034053-90f777

661,8 tok/s

3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260727-193100-9a2d59

615,3 tok/s

3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-032119-325be2

614,4 tok/s

3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-032119-7e2418

606,1 tok/s

3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-032119-f04db0

530,9 tok/s

2× NVIDIA RTX A6000run-20260830-061630-a3022b

276,5 tok/s

2× NVIDIA RTX A6000run-20260827-224816-0f7cf8

269,5 tok/s

3× AMD Radeon AI PRO R9700run-20260723-045450-a75a33

104,1 tok/s

2× AMD Radeon PRO W7900 Dual Slotrun-20260818-073335-52522e

99,5 tok/s

AMD Radeon PRO W7900 Dual Slotrun-20260817-171106-73efdf

29,7 tok/s

NVIDIA GeForce RTX 3090 Tirun-20260727-202226-b7012c

20,8 tok/s

NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition this runrun-20260924-131556-ded9b8

18,8 tok/s

2× AMD Radeon PRO W7900 Dual Slotrun-20260818-083746-9c0b64

9,0 tok/s

NVIDIA GeForce RTX 5070 Tirun-20260727-142111-faf33c

2,6 tok/s

Hardware

GPU: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB VRAM
CPU: AMD Ryzen 9 7945HX with Radeon Graphics
RAM: 60 GB
Mainboard: Shenzhen Meigao Electronic Equipment Co.,Ltd DRFXI (MotherBoard Series)

Setup

Runtime: vLLM
Quantization: FP8
Model: GLM-4.5-Air

Configuration

benchmark-konfiguration — run-20260924-131556-ded9b8

# LLM-Benchmark Konfiguration # Modell : GLM-4.5-Air # Engine : vLLM # Run-ID : run-20260924-131556-ded9b8 # GPU : NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition # CPU : AMD Ryzen 9 7945HX with Radeon Graphics # RAM : 60 GB bench@llm-benchmark:~$ docker vllm/vllm-openai serve GLM-4.5-Air-FP8 \ --served-model-name GLM-4.5-Air \ --gpu-memory-utilization 0.90 \ --max-model-len 16384 \ --max-num-seqs 12 \ --cpu-offload-gb 32 '(NVIDIA' RTX PRO 6000 Blackwell '96GB)'

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.	vllm
Modellalias?Der Name, unter dem das Modell ueber die API angesprochen wird. Genau dieser Wert muss im Request-Feld 'model' stehen.	GLM-4.5-Air
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.	16384
Alias?Anzeigename des Modells nach aussen (served model name), unabhaengig vom Dateinamen.	GLM-4.5-Air
GPU-Speicher?Anteil des GPU-Speichers (0 bis 1), den vLLM belegen darf. 0.92 = 92 %. Hoeher = mehr Platz fuer den KV-Cache (mehr/laengere parallele Anfragen), aber groesseres Risiko fuer 'Out of Memory'.	0.90
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.	16384
Max-Sequenzen?Maximale Anzahl gleichzeitig bearbeiteter Anfragen (Batch-Groesse). Hoeher = mehr Durchsatz, braucht aber mehr KV-Cache-Speicher.	12
cpu-offload-gb?Menge an Modellgewichten in GiB, die in den CPU-RAM ausgelagert wird. Ermoeglicht groessere Modelle als der GPU-Speicher fasst, kostet aber Geschwindigkeit.	32

All benchmarks of this model To leaderboard

Model comparison

GLM-4.5-Air 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 RTX PRO 6000 Blackwell Workstation Edition 661,8 tok/s★ NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 615,3 tok/s this runNVIDIA RTX A6000 276,5 tok/sAMD Radeon AI PRO R9700 104,1 tok/sAMD Radeon PRO W7900 Dual Slot 99,5 tok/sNVIDIA GeForce RTX 3090 Ti 20,8 tok/sNVIDIA GeForce RTX 5070 Ti 9,9 tok/sNVIDIA GeForce RTX 5090 6,2 tok/sNVIDIA GeForce RTX 2060 4,2 tok/sCPU-only 2,1 tok/s

CPUby processor

AMD Ryzen 9 9950X 16-Core Processor 661,8 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 615,3 tok/sAMD Ryzen Threadripper PRO 7955WX 16-Cores 276,5 tok/sAMD Ryzen Threadripper PRO 5975WX 32-Cores 99,5 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 20,8 tok/s★ AMD Ryzen 9 7945HX with Radeon Graphics 18,8 tok/s this runAMD Ryzen 7 5800X3D 8-Core Processor 6,2 tok/sIntel(R) Xeon(R) CPU E5-4657L v2 @ 2.40GHz 4,2 tok/sIntel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz 2,1 tok/s

MBby mainboard

ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 661,8 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 615,3 tok/sASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 99,5 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 20,8 tok/s★ Shenzhen Meigao Electronic Equipment Co.,Ltd DRFXI (MotherBoard Series) 18,8 tok/s this runASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 6,2 tok/sDell Inc. PowerEdge R820 4,2 tok/s

ENGby engine

llama.cpp 661,8 tok/s★ vLLM 276,5 tok/s this rununbekannt 32,9 tok/s

DRVby driver

unbekannt 661,8 tok/sAMD 7.0.0-27-generic 32,9 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 URL Reset

⚡ Electricity price EUR/kWh

⚙️ System utilization 100 %

🖥️ Acquisition EUR

🔌 Idle 30 W

⚡ TDP 300 W

☁️ External LLM (API) show

Input €/1M Output €/1M Input per output token

Electricity0.30 EUR/kWh

Avg power (incl. idle)300 W estimated (TDP)GPU 300 W full load

Avg cost / hourEUR 0.090

Electricity / 1M tokensEUR 1.33

Token / kWh225.00K

Acquisition (system)EUR 13,120 partial priceGPU EUR 13,000 · PSU EUR 120

Electricity (2 years)–

TCO (2 years)EUR 14,697

Output tokens (2 years)1.18B

☁️ 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 (30 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

	GLM-4.5-AirNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition	GLM-4.5-AirNVIDIA RTX PRO 6000 Blackwell Workstation Edition	GLM-4.5-Air3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition	GLM-4.5-Air3x 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 RTX PRO 6000 Blackwell Max-Q Workstation Edition · this run 19 tok/s NVIDIA RTX PRO 6000 Blackwell Workstation Edition 662 tok/s 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 615 tok/s 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 614 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.