Created bymario-alka.dePowered bygodcore.denoob2claw.detricoma.de
Contributed byMario AlkaNVIDIA

Nemotron-3-Super-120B-A12B

Performance benchmark · measured on 17.08.2026 17:08

Benchmark-IDrun-20260817-171106-934436
Timebench 3 - Kombi (Prefill + Generation)MoE120BRuntime: llama.cppQuantisierung: Q4_K_M
Generation29,17tok/s
Prefill851,46tok/s
Time to First Token2.527,50ms
Total duration75,27s
Concurrency1parallel
Ranking in the field
215of 455 systems

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

This run is better than 53 % of all comparable systems.
Generation 29,2 tok/s
-17 % vs Ø 35,3
Prefill 851,5 tok/s
-57 % vs Ø 1.966,4
Time to First Token 2.528 ms
-93 % vs Ø 35.848
Distribution in the field1 – 148 tok/s
Ø 35 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: 4x AMD Radeon AI PRO R9700 · 32 GB VRAM
CPU: AMD Ryzen Threadripper PRO 7955WX 16-Cores
RAM: 184 GB
Mainboard: ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE

Setup

Runtime: llama.cpp
Quantization: Q4_K_M
Model: Nemotron-3-Super-120B-A12B

Configuration

benchmark-konfiguration — run-20260817-171106-934436
# LLM-Benchmark Konfiguration # Modell : Nemotron-3-Super-120B-A12B # Engine : llama.cpp # Run-ID : run-20260817-171106-934436 # GPU : 4x AMD Radeon AI PRO R9700 # CPU : AMD Ryzen Threadripper PRO 7955WX 16-Cores # RAM : 184 GB bench@llm-benchmark:~$ /home/godcore/llama.cpp/build/bin/llama-server \ -m /home/godcore/.cache/huggingface/hub/models--DevQuasar--nvidia.NVIDIA-Nemotron-3-Super-120B-A12B-BF16-GGUF/snapshots/fe70b8fa41864e382a726d44b138588c124d7337/nvidia.NVIDIA-Nemotron-3-Super-120B-A12B-BF16.Q4_K_M-00001-of-00006.gguf \ --alias NVIDIA-Nemotron-3-Super-120B-A12B \ -ngl 999 \ -fa on \ -sm layer \ -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.NVIDIA-Nemotron-3-Super-120B-A12B
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/.cache/huggingface/hub/models--DevQuasar--nvidia.NVIDIA-Nemotron-3-Super-120B-A12B-BF16-GGUF/snapshots/fe70b8fa41864e382a726d44b138588c124d7337/nvidia.NVIDIA-Nemotron-3-Super-120B-A12B-BF16.Q4_K_M-00001-of-00006.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
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

Nemotron-3-Super-120B-A12B 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

7805853901950,009641.9272.891Prefill (tok/s)Generation (tok/s)NVIDIA RTX PRO 6000 Blackwell Workstation Edition - 600,9 tok/s Generation, 837 tok/s Prefill, TTFT 73.953 ms (9 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition - 423,4 tok/s Generation, 2.339 tok/s Prefill, TTFT 12.569 ms (12 Laufe)NVIDIA RTX PRO 6000 B...NVIDIA GeForce RTX 5070 Ti - 32,6 tok/s Generation, 88 tok/s Prefill, TTFT 156.105 ms (6 Laufe)NVIDIA GeForce RTX 50...NVIDIA GeForce RTX 3090 Ti - 18,9 tok/s Generation, 106 tok/s Prefill, TTFT 171.396 ms (3 Laufe)NVIDIA GeForce RTX 30...NVIDIA GB10 (DGX Spark) - 11,4 tok/s Generation, 1.495 tok/s Prefill, TTFT 1.449 ms (1 Lauf)NVIDIA GB10 (DGX Spar...NVIDIA GeForce RTX 5090 - 5,5 tok/s Generation, 49 tok/s Prefill, TTFT 150.808 ms (8 Laufe)NVIDIA GeForce RTX 50...AMD Radeon AI PRO R9700 - 80,5 tok/s Generation, 436 tok/s Prefill, TTFT 102.716 ms (17 Laufe) | DIESER LAUF★ AMD Radeon AI PRO R97...
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 600,9 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition 423,4 tok/s★ AMD Radeon AI PRO R9700 80,5 tok/s this runNVIDIA GeForce RTX 5070 Ti 32,6 tok/sNVIDIA GeForce RTX 3090 Ti 18,9 tok/sNVIDIA GB10 (DGX Spark) 11,4 tok/sNVIDIA GeForce RTX 5090 5,5 tok/s

CPUby processor

7805853901950,009641.9272.891Prefill (tok/s)Generation (tok/s)AMD Ryzen 9 9950X 16-Core Processor - 600,9 tok/s Generation, 837 tok/s Prefill, TTFT 73.953 ms (9 Laufe)AMD Ryzen 9 9950X 16-...AMD Ryzen Threadripper PRO 9965WX 24-Cores - 423,4 tok/s Generation, 2.339 tok/s Prefill, TTFT 12.569 ms (12 Laufe)AMD Ryzen Threadrippe...AMD Ryzen Threadripper PRO 5975WX 32-Cores - 32,6 tok/s Generation, 88 tok/s Prefill, TTFT 156.105 ms (6 Laufe)AMD Ryzen Threadrippe...AMD Ryzen 9 8945HX with Radeon Graphics - 18,9 tok/s Generation, 106 tok/s Prefill, TTFT 171.396 ms (3 Laufe)AMD Ryzen 9 8945HX wi...NVIDIA Grace - 11,4 tok/s Generation, 1.495 tok/s Prefill, TTFT 1.449 ms (1 Lauf)NVIDIA GraceAMD Ryzen 7 5800X3D 8-Core Processor - 5,5 tok/s Generation, 49 tok/s Prefill, TTFT 150.808 ms (8 Laufe)AMD Ryzen 7 5800X3D 8...AMD Ryzen Threadripper PRO 7955WX 16-Cores - 80,5 tok/s Generation, 436 tok/s Prefill, TTFT 102.716 ms (17 Laufe) | DIESER LAUF★ AMD Ryzen Threadrippe...
AMD Ryzen 9 9950X 16-Core Processor 600,9 tok/sAMD Ryzen Threadripper PRO 9965WX 24-Cores 423,4 tok/s★ AMD Ryzen Threadripper PRO 7955WX 16-Cores 80,5 tok/s this runAMD Ryzen Threadripper PRO 5975WX 32-Cores 32,6 tok/sAMD Ryzen 9 8945HX with Radeon Graphics 18,9 tok/sNVIDIA Grace 11,4 tok/sAMD Ryzen 7 5800X3D 8-Core Processor 5,5 tok/s

MBby mainboard

7805853901950,006151.2301.845Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI - 600,9 tok/s Generation, 837 tok/s Prefill, TTFT 73.953 ms (9 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI - 32,6 tok/s Generation, 88 tok/s Prefill, TTFT 156.105 ms (6 Laufe)ASUSTeK COMPUTER INC....Meigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) - 18,9 tok/s Generation, 106 tok/s Prefill, TTFT 171.396 ms (3 Laufe)Meigao Innovation Tec...ASUSTeK COMPUTER INC. GX10 - 11,4 tok/s Generation, 1.495 tok/s Prefill, TTFT 1.449 ms (1 Lauf)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING - 5,5 tok/s Generation, 49 tok/s Prefill, TTFT 150.808 ms (8 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE - 423,4 tok/s Generation, 1.223 tok/s Prefill, TTFT 65.414 ms (29 Laufe) | DIESER LAUF★ ASUSTeK COMPUTER INC....
ASUSTeK COMPUTER INC. ProArt X870E-CREATOR WIFI 600,9 tok/s★ ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 423,4 tok/s this runASUSTeK COMPUTER INC. Pro WS WRX80E-SAGE SE WIFI 32,6 tok/sMeigao Innovation Technology (Shen Zhen) Co., Ltd DRFXL (MotherBoard Series) 18,9 tok/sASUSTeK COMPUTER INC. GX10 11,4 tok/sASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 5,5 tok/s

ENGby engine

7795843891950,05909631.3361.709Prefill (tok/s)Generation (tok/s)vLLM - 11,4 tok/s Generation, 1.495 tok/s Prefill, TTFT 1.449 ms (1 Lauf)vLLMllama.cpp - 600,9 tok/s Generation, 804 tok/s Prefill, TTFT 94.906 ms (55 Laufe) | DIESER LAUF★ llama.cpp
★ llama.cpp 600,9 tok/s this runvLLM 11,4 tok/s

DRVby driver

7795843891950,05909631.3361.709Prefill (tok/s)Generation (tok/s)unbekannt - 600,9 tok/s Generation, 804 tok/s Prefill, TTFT 94.906 ms (55 Laufe)unbekanntNVIDIA 590.48.01 / CUDA 13.1 - 11,4 tok/s Generation, 1.495 tok/s Prefill, TTFT 1.449 ms (1 Lauf)NVIDIA 590.48.01 / CU...
unbekannt 600,9 tok/sNVIDIA 590.48.01 / CUDA 13.1 11,4 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 145 W
⚡ TDP 1271 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)1,271 W estimated (TDP)GPU 1,200 + CPU 46 + Board 25 W full load
Avg cost / hourEUR 0.38
Electricity / 1M tokensEUR 3.63
Token / kWh82.62K
Acquisition (system)EUR 11,114 full priceGPU EUR 5,600 · CPU EUR 1,399 · Board EUR 1,299 · RAM EUR 2,576 · PSU EUR 240
Electricity (2 years)
TCO (2 years)EUR 17,794
Output tokens (2 years)1.84B
☁️ 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 (145 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-3-Super-120B-A12B4x AMD Radeon AI PRO R9700Nemotron-3-Super-120B-A12BNVIDIA RTX PRO 6000 Blackwell Workstation EditionNemotron-3-Super-120B-A12B3x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNemotron-3-Super-120B-A12B3x 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.