Created bymario-alka.dePowered bygodcore.denoob2claw.detricoma.de
Contributed byMario AlkaZ.ai (Zhipu)

GLM-4.5-Air

Performance benchmark · measured on 18.08.2026 07:44

Benchmark-IDrun-20260818-080548-ece3de
Timebench 3 - Kombi (Prefill + Generation)MoE106BRuntime: llama.cppQuantisierung: Q8_0
Generation11,22tok/s
Prefill102,83tok/s
Time to First Token20.742,00ms
Total duration224,10s
Concurrency1parallel
Ranking in the field
876of 1161 systems

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

This run is better than 25 % of all comparable systems.
Generation 11,2 tok/s
-85 % vs Ø 74,3
Prefill 102,8 tok/s
-96 % vs Ø 2.324,7
Time to First Token 20.742 ms
-42 % vs Ø 35.657
Distribution in the field0 – 405 tok/s
Ø 74 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 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: Q8_0
Model: GLM-4.5-Air

Configuration

benchmark-konfiguration — run-20260818-080548-ece3de
# LLM-Benchmark Konfiguration # Modell : GLM-4.5-Air # Engine : llama.cpp # Run-ID : run-20260818-080548-ece3de # 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/zai-org_GLM-4.5-Air-Q8_0-00001-of-00003.gguf \ --alias GLM-4.5-Air \ -ngl 999 \ -fa on \ -sm layer \ --tensor-split 1,1 \ -cmoe \ -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.GLM-4.5-Air
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/zai-org_GLM-4.5-Air-Q8_0-00001-of-00003.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
faon
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
np12

All benchmarks of this model To leaderboard

Anzeige
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

8606454302150,001.4242.8484.272Prefill (tok/s)Generation (tok/s)NVIDIA RTX PRO 6000 Blackwell Workstation Edition - 661,8 tok/s Generation, 3.117 tok/s Prefill, TTFT 8.401 ms (3 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition - 615,3 tok/s Generation, 3.448 tok/s Prefill, TTFT 9.105 ms (12 Laufe)NVIDIA RTX PRO 6000 B...AMD Radeon AI PRO R9700 - 104,1 tok/s Generation, 1.403 tok/s Prefill, TTFT 67.488 ms (11 Laufe)AMD Radeon AI PRO R97...NVIDIA GeForce RTX 3090 Ti - 20,8 tok/s Generation, 137 tok/s Prefill, TTFT 156.298 ms (3 Laufe)NVIDIA GeForce RTX 30...NVIDIA GeForce RTX 5070 Ti - 9,9 tok/s Generation, 109 tok/s Prefill, TTFT 59.573 ms (3 Laufe)NVIDIA GeForce RTX 50...NVIDIA GeForce RTX 5090 - 6,2 tok/s Generation, 161 tok/s Prefill, TTFT 57.809 ms (5 Laufe)NVIDIA GeForce RTX 50...CPU-only - 2,1 tok/s Generation, 18 tok/s Prefill, TTFT 117.963 ms (1 Lauf)CPU-onlyAMD Radeon PRO W7900 Dual Slot - 99,5 tok/s Generation, 461 tok/s Prefill, TTFT 57.896 ms (9 Laufe) | DIESER LAUF★ AMD Radeon PRO W7900 ...
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 661,8 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 615,3 tok/sAMD Radeon AI PRO R9700 104,1 tok/s★ AMD Radeon PRO W7900 Dual Slot 99,5 tok/s this runNVIDIA GeForce RTX 3090 Ti 20,8 tok/sNVIDIA GeForce RTX 5070 Ti 9,9 tok/sNVIDIA GeForce RTX 5090 6,2 tok/sCPU-only 2,1 tok/s

CPUby processor

8606454302150,001.4242.8484.272Prefill (tok/s)Generation (tok/s)AMD Ryzen 9 9950X 16-Core Processor - 661,8 tok/s Generation, 3.117 tok/s Prefill, TTFT 8.401 ms (3 Laufe)AMD Ryzen 9 9950X 16-...AMD Ryzen Threadripper PRO 9965WX 24-Cores - 615,3 tok/s Generation, 3.448 tok/s Prefill, TTFT 9.105 ms (12 Laufe)AMD Ryzen Threadrippe...AMD Ryzen Threadripper PRO 7955WX 16-Cores - 104,1 tok/s Generation, 1.403 tok/s Prefill, TTFT 67.488 ms (11 Laufe)AMD Ryzen Threadrippe...AMD Ryzen 9 8945HX with Radeon Graphics - 20,8 tok/s Generation, 137 tok/s Prefill, TTFT 156.298 ms (3 Laufe)AMD Ryzen 9 8945HX wi...AMD Ryzen 7 5800X3D 8-Core Processor - 6,2 tok/s Generation, 161 tok/s Prefill, TTFT 57.809 ms (5 Laufe)AMD Ryzen 7 5800X3D 8...Intel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz - 2,1 tok/s Generation, 18 tok/s Prefill, TTFT 117.963 ms (1 Lauf)Intel(R) Xeon(R) CPU ...AMD Ryzen Threadripper PRO 5975WX 32-Cores - 99,5 tok/s Generation, 373 tok/s Prefill, TTFT 58.315 ms (12 Laufe) | DIESER LAUF★ AMD Ryzen Threadrippe...
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 104,1 tok/s★ AMD Ryzen Threadripper PRO 5975WX 32-Cores 99,5 tok/s this runAMD Ryzen 9 8945HX with Radeon Graphics 20,8 tok/sAMD Ryzen 7 5800X3D 8-Core Processor 6,2 tok/sIntel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz 2,1 tok/s

MBby mainboard

8606454302150,001.2872.5753.862Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI - 661,8 tok/s Generation, 3.117 tok/s Prefill, TTFT 8.401 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE - 615,3 tok/s Generation, 2.470 tok/s Prefill, TTFT 37.028 ms (23 Laufe)ASUSTeK COMPUTER INC....Meigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) - 20,8 tok/s Generation, 137 tok/s Prefill, TTFT 156.298 ms (3 Laufe)Meigao Innovation Tec...ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING - 6,2 tok/s Generation, 161 tok/s Prefill, TTFT 57.809 ms (5 Laufe)ASUSTeK COMPUTER INC....Dell Inc. PowerEdge R820 - 2,1 tok/s Generation, 18 tok/s Prefill, TTFT 117.963 ms (1 Lauf)Dell Inc. PowerEdge R...ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 99,5 tok/s Generation, 373 tok/s Prefill, TTFT 58.315 ms (12 Laufe) | DIESER LAUF★ ASUSTeK COMPUTER INC....
ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 661,8 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 615,3 tok/s★ ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 99,5 tok/s this runMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 20,8 tok/sASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 6,2 tok/sDell Inc. PowerEdge R820 2,1 tok/s

ENGby engine

8596444292150,09351.8232.7113.600Prefill (tok/s)Generation (tok/s)unbekannt - 32,9 tok/s Generation, 3.110 tok/s Prefill, TTFT 1.893 ms (2 Laufe)unbekanntvLLM - 9,1 tok/s Generation, 1.934 tok/s Prefill, TTFT 15.924 ms (3 Laufe)vLLMllama.cpp - 661,8 tok/s Generation, 1.425 tok/s Prefill, TTFT 57.166 ms (42 Laufe) | DIESER LAUF★ llama.cpp
★ llama.cpp 661,8 tok/s this rununbekannt 32,9 tok/svLLM 9,1 tok/s

DRVby driver

8546404272130,09751.8482.7213.593Prefill (tok/s)Generation (tok/s)unbekannt - 661,8 tok/s Generation, 1.459 tok/s Prefill, TTFT 54.416 ms (45 Laufe)unbekanntAMD 7.0.0-27-generic - 32,9 tok/s Generation, 3.110 tok/s Prefill, TTFT 1.893 ms (2 Laufe)AMD 7.0.0-27-generic
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 (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 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 4.46
Token / kWh67.32K
Acquisition (system)EUR 2,156 missingRAM EUR 1,976 · PSU EUR 180
Electricity (2 years)
TCO (2 years)EUR 5,310
Output tokens (2 years)707.67M
☁️ 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

GLM-4.5-Air2x AMD Radeon PRO W7900 Dual SlotGLM-4.5-AirNVIDIA RTX PRO 6000 Blackwell Workstation EditionGLM-4.5-Air3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionGLM-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)

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.