Contributed byMario AlkaNVIDIA Nemotron-Cascade-2-30B-A3B
Performance benchmark · measured on 20.08.2026 12:45
Benchmark-ID
run-20260820-131510-2cf63bTimebench 3 - Kombi (Prefill + Generation)MoE30BRuntime: llama.cppQuantisierung: Q8_0
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.
Generation101,36tok/s
Prefill3.237,32tok/s
Time to First Token6.849,50ms
Total duration216,40s
Concurrency10parallel
Ranking in the field
21of 24 systems
Performance benchmark · Primary metric: Generation-Speed (tok/s) · 10× concurrent
This run is better than 13 % of all comparable systems.
Generation
101,4 tok/s
-62 % vs Ø 269,0
Prefill
3.237,3 tok/s
-28 % vs Ø 4.519,5
Time to First Token
6.850 ms
-68 % vs Ø 21.234
Distribution in the field64 – 903 tok/s
Wie schlägt sich dieser Benchmark mit anderen Modellen?
gemma-4-E2B-itNVIDIA GB10 · run-20260729-224611-cc3776
903,3 tok/s
gemma-4-E2B-itNVIDIA GB10 · run-20260729-224610-32ded2
860,9 tok/s
gemma-4-E4B-itNVIDIA GB10 · run-20260730-084906-3e426d
488,2 tok/s
gemma-4-E4B-itNVIDIA GB10 · run-20260729-224611-9fe5af
484,8 tok/s
gemma-4-E4B-itNVIDIA GB10 · run-20260729-224611-0a62d6
455,1 tok/s
Nemotron-3-Nano-4BNVIDIA GB10 · run-20260730-084906-84663c
367,8 tok/s
Nemotron-3-Nano-4BNVIDIA GB10 · run-20260730-035052-93e207
354,3 tok/s
Ling-3.0-flashNVIDIA GB10 · run-20260820-173706-52a2df
253,7 tok/s
Ling-3.0-flashNVIDIA GB10 · run-20260820-131510-16a6a0
250,7 tok/s
Ling-3.0-flashNVIDIA GB10 · run-20260820-131510-6c0f64
247,8 tok/s
gpt-oss-20bNVIDIA GB10 · run-20260820-194812-1bb2eb
202,1 tok/s
Qwen3-Coder-30B-A3B-InstructNVIDIA GB10 · run-20260820-205402-b9c794
182,9 tok/s
MiniMax-M2.7NVIDIA GB10 · run-20260723-055735-2facb9
174,2 tok/s
Qwen3-30B-A3B-Instruct-2507NVIDIA GB10 · run-20260820-194812-a67646
169,9 tok/s
Nemotron-Cascade-2-30B-A3B this runNVIDIA GB10 · run-20260820-131510-2cf63b
101,4 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-20260729-032120-1b1a49
1.768,1 tok/s
NVIDIA GeForce RTX 5090run-20260729-105335-34f830
1.719,2 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-105340-3fa079
1.655,2 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-105340-b8a790
1.650,7 tok/s
3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Editionrun-20260729-105340-54128f
1.542,6 tok/s
NVIDIA GeForce RTX 3090 Tirun-20260729-060802-b93c1e
897,5 tok/s
NVIDIA GeForce RTX 5070 Tirun-20260729-032126-39e622
262,4 tok/s
NVIDIA GeForce RTX 5070 Tirun-20260818-080548-666d0c
219,6 tok/s
NVIDIA GeForce RTX 5070 Tirun-20260818-071253-1a8280
190,3 tok/s
NVIDIA GB10run-20260820-124207-6d1c5d
148,8 tok/s
3× AMD Radeon AI PRO R9700run-20260806-094441-fc7051
122,1 tok/s
3× AMD Radeon AI PRO R9700run-20260806-094441-fb0970
122,1 tok/s
3× AMD Radeon AI PRO R9700run-20260806-094441-8c3f52
121,3 tok/s
NVIDIA GB10 this runrun-20260820-131510-2cf63b
101,4 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: Q8_0
Driver: NVIDIA 590.48.01 / CUDA 13.1
Model: Nemotron-Cascade-2-30B-A3B
Configuration
# LLM-Benchmark Konfiguration
# Modell : Nemotron-Cascade-2-30B-A3B
# Engine : llama.cpp
# Run-ID : run-20260820-131510-2cf63b
# GPU : NVIDIA GB10
# CPU : NVIDIA Grace
# RAM : 120 GB
bench@llm-benchmark:~$ /home/godcore/llama.cpp/build/bin/llama-server \
-m /home/godcore/nemc2/nvidia_Nemotron-Cascade-2-30B-A3B-Q8_0.gguf \
--alias Nemotron-Cascade-2-30B-A3B \
-ngl 999 \
-fa on \
-c 49152 \
-np 12 \
-b 8192 \
-ub 4096 \
--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. | Nemotron-Cascade-2-30B-A3B |
| 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/nemc2/nvidia_Nemotron-Cascade-2-30B-A3B-Q8_0.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 |
| Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells. | 49152 |
| np | 12 |
| Batch-Groesse?Logische Batchgroesse fuer die Prompt-Verarbeitung (Anzahl Token pro Prefill-Durchlauf). | 8192 |
| Micro-Batch?Physische Micro-Batchgroesse. Begrenzt, wie viele Token gleichzeitig durch die Compute-Pipeline laufen. | 4096 |
Anzeige
Model comparison
Nemotron-Cascade-2-30B-A3B 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 1.768,1 tok/sNVIDIA GeForce RTX 5090 1.719,2 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1.655,2 tok/sNVIDIA GeForce RTX 3090 Ti 897,5 tok/sAMD Radeon 8060S Graphics 326,9 tok/sNVIDIA GeForce RTX 5070 Ti 262,4 tok/s★ NVIDIA GB10 (DGX Spark) 148,8 tok/s this runAMD Radeon AI PRO R9700 122,1 tok/sCPU-only 5,9 tok/s
CPUby processor
AMD Ryzen 9 9950X 16-Core Processor 1.768,1 tok/sAMD Ryzen 7 5800X3D 8-Core Processor 1.719,2 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 1.655,2 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 897,5 tok/sAMD RYZEN AI MAX+ 395 w/ Radeon 8060S 326,9 tok/sAMD Ryzen Threadripper PRO 5975WX 32-Cores 262,4 tok/sAMD Ryzen 9 7945HX with Radeon Graphics 219,6 tok/s★ NVIDIA Grace 148,8 tok/s this runAMD Ryzen Threadripper PRO 7955WX 16-Cores 122,1 tok/sIntel(R) Xeon(R) CPU E5-4620 v2 @ 2.60GHz 5,9 tok/s
MBby mainboard
ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 1.768,1 tok/sASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 1.719,2 tok/sASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 1.655,2 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 897,5 tok/sBosgame AXB35-02 (BeyondMax Series) 326,9 tok/sASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 262,4 tok/sShenzhen Meigao Electronic Equipment Co.,Ltd DRFXI (MotherBoard Series) 219,6 tok/s★ ASUSTeK COMPUTER INC. GX10 148,8 tok/s this runDell Inc. PowerEdge R820 5,9 tok/s
DRVby driver
unbekannt 1.768,1 tok/s★ NVIDIA 590.48.01 / CUDA 13.1 148,8 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 (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
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 0.12
Token / kWh2.61M
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)6.39B
☁️ 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
| Nemotron-Cascade-2-30B-A3BNVIDIA GB10 (DGX Spark) | Nemotron-Cascade-2-30B-A3BNVIDIA RTX PRO 6000 Blackwell Workstation Edition | Nemotron-Cascade-2-30B-A3BNVIDIA GeForce RTX 5090 | Nemotron-Cascade-2-30B-A3B3x 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 101 tok/sNVIDIA RTX PRO 6000 Blackwell Workstation Edition 1,768 tok/sNVIDIA GeForce RTX 5090 1,719 tok/s3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 1,655 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.