Created bymario-alka.dePowered bynoob2claw.detricoma.de
Contributed byMarcel SommerNVIDIA

NVIDIA-Nemotron-3.5-Lightning-30B-A3B

Performance benchmark · measured on 12.08.2026 12:35

Benchmark-IDrun-20261021-093815-b22926
Timebench 3 - Kombi (Prefill + Generation)MoE31.6BRuntime: vLLMQuantisierung: NVFP4
Generation1.134,97tok/s
Prefill8.756,90tok/s
Time to First Token9.812,00ms
Total duration47,74s
Concurrency5parallel
Ranking in the field
12of 1539 systems

Performance benchmark · Primary metric: Generation-Speed (tok/s) · 5× concurrent

This run is better than 99 % of all comparable systems.
Generation 1.135,0 tok/s
+339 % vs Ø 258,5
Prefill 8.756,9 tok/s
+59 % vs Ø 5.498,6
Time to First Token 9.812 ms
-66 % vs Ø 28.923
Distribution in the field0 – 1.695 tok/s
Ø 258 Median Dieser Lauf

Wie schlägt sich dieser Benchmark mit anderen Modellen?

How does this benchmark compare on other GPUs?

Same model on different hardware · 5× concurrent · Generation (tok/s)

Hardware

GPU: NVIDIA GeForce RTX 3090 Ti · 24 GB VRAM
CPU: AMD Ryzen 5 5600X 6-Core Processor
RAM: 30 GB
Mainboard: ASUSTeK COMPUTER INC. PRIME A520M-K

Setup

Runtime: vLLM
Quantization: NVFP4
Model: NVIDIA-Nemotron-3.5-Lightning-30B-A3B

Configuration

benchmark-konfiguration — run-20261021-093815-b22926
# LLM-Benchmark Konfiguration # Modell : NVIDIA-Nemotron-3.5-Lightning-30B-A3B # Engine : vLLM # Run-ID : run-20261021-093815-b22926 # GPU : NVIDIA GeForce RTX 3090 Ti # CPU : AMD Ryzen 5 5600X 6-Core Processor # RAM : 30 GB bench@llm-benchmark:~$ /opt/vllm-nemotron/venv/bin/python /opt/vllm-nemotron/venv/bin/vllm serve /home/godcore/models/nemotron35-lightning-30b-a3b-nvfp4 \ --served-model-name nemotron-3.5-lightning-30b-a3b \ --host 192.168.41.116 \ --port 8000 \ --max-model-len 8192 \ --max-num-seqs 1 \ --max-num-batched-tokens 8192 \ --gpu-memory-utilization 0.92 \ --quantization modelopt_fp4 \ --moe-backend humming \ --linear-backend humming \ --mamba-backend flashinfer \ --mamba-cache-mode align \ --mamba-ssu-algorithm simple \ --reasoning-parser nemotron_v3 \ --trust-remote-code
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.nemotron-3.5-lightning-30b-a3b
Kontextlaenge?Maximale Anzahl Tokens (Eingabe + erzeugte Ausgabe zusammen), die das Modell pro Anfrage verarbeiten kann.8192
Alias?Anzeigename des Modells nach aussen (served model name), unabhaengig vom Dateinamen.nemotron-3.5-lightning-30b-a3b
Kontext?Groesse des Kontextfensters in Token. 0 = der beim Training verwendete Kontext des Modells.8192
Max-Sequenzen?Maximale Anzahl gleichzeitig bearbeiteter Anfragen (Batch-Groesse). Hoeher = mehr Durchsatz, braucht aber mehr KV-Cache-Speicher.1
Max-Batch-Tokens?Maximale Anzahl Tokens, die pro Rechenschritt gebuendelt verarbeitet werden. Beeinflusst Prefill-Durchsatz und Latenz.8192
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'.0.92
Quantisierung?Quantisierungsverfahren der Gewichte (z.B. awq, gptq, fp8, bitsandbytes). Verkleinert das Modell und spart VRAM, kann die Genauigkeit leicht senken. Leer = keine zusaetzliche Quantisierung.modelopt_fp4
moe-backendhumming
linear-backendhumming
mamba-backendflashinfer
mamba-cache-modealign
mamba-ssu-algorithmsimple
Reasoning-Parser?Trennt den Denk-/Reasoning-Teil der Antwort vom eigentlichen Inhalt (z.B. bei Modellen mit <think>-Bloecken).nemotron_v3
Trust-Remote-Code?Erlaubt vLLM, mit dem Modell mitgelieferten Python-Code auszufuehren. Noetig fuer Architekturen, die vLLM nicht von Haus aus kennt. Nur bei vertrauenswuerdigen Modellen aktivieren.aktiv
execution_typelocal

All benchmarks of this model To leaderboard

Anzeige
Model comparison

NVIDIA-Nemotron-3.5-Lightning-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

4.5023.3772.2511.1260,004.6219.24313.864Prefill (tok/s)Generation (tok/s)NVIDIA GeForce RTX 3090 - 3.484,3 tok/s Generation, 11.447 tok/s Prefill, TTFT 3.078 ms (29 Laufe)NVIDIA GeForce RTX 30...AMD Radeon AI PRO R9700 - 290,3 tok/s Generation, 6.285 tok/s Prefill, TTFT 963 ms (30 Laufe)AMD Radeon AI PRO R97...NVIDIA RTX SPARK (N1X) - 146,9 tok/s Generation, 1.835 tok/s Prefill, TTFT 5.732 ms (4 Laufe)NVIDIA RTX SPARK (N1X)NVIDIA DGX SPARK (GB10) - 136,6 tok/s Generation, 2.246 tok/s Prefill, TTFT 5.100 ms (3 Laufe)NVIDIA DGX SPARK (GB1...NVIDIA GeForce RTX 3090 Ti - 2.263,7 tok/s Generation, 8.103 tok/s Prefill, TTFT 10.659 ms (3 Laufe) | DIESER LAUF★ NVIDIA GeForce RTX 30...
NVIDIA GeForce RTX 3090 3.484,3 tok/s★ NVIDIA GeForce RTX 3090 Ti 2.263,7 tok/s this runAMD Radeon AI PRO R9700 290,3 tok/sNVIDIA RTX SPARK (N1X) 146,9 tok/sNVIDIA DGX SPARK (GB10) 136,6 tok/s

CPUby processor

4.5023.3772.2511.1260,004.4838.96613.450Prefill (tok/s)Generation (tok/s)AMD Ryzen Threadripper PRO 7955WX 16-Cores - 290,3 tok/s Generation, 6.619 tok/s Prefill, TTFT 679 ms (27 Laufe)AMD Ryzen Threadrippe...AMD Ryzen 3 3100 4-Core Processor - 283,5 tok/s Generation, 3.276 tok/s Prefill, TTFT 3.524 ms (3 Laufe)AMD Ryzen 3 3100 4-Co...NVIDIA RTX Spark N1X (6144-core GPU, 20-core CPU) - 146,9 tok/s Generation, 1.835 tok/s Prefill, TTFT 5.732 ms (4 Laufe)NVIDIA RTX Spark N1X ...NVIDIA Grace - 136,6 tok/s Generation, 2.246 tok/s Prefill, TTFT 5.100 ms (3 Laufe)NVIDIA GraceAMD Ryzen 5 5600X 6-Core Processor - 3.484,3 tok/s Generation, 11.113 tok/s Prefill, TTFT 3.789 ms (32 Laufe) | DIESER LAUF★ AMD Ryzen 5 5600X 6-C...
★ AMD Ryzen 5 5600X 6-Core Processor 3.484,3 tok/s this runAMD Ryzen Threadripper PRO 7955WX 16-Cores 290,3 tok/sAMD Ryzen 3 3100 4-Core Processor 283,5 tok/sNVIDIA RTX Spark N1X (6144-core GPU, 20-core CPU) 146,9 tok/sNVIDIA Grace 136,6 tok/s

MBby mainboard

4.5023.3772.2511.1260,004.6219.24313.864Prefill (tok/s)Generation (tok/s)ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING - 3.484,3 tok/s Generation, 11.447 tok/s Prefill, TTFT 3.078 ms (29 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE - 290,3 tok/s Generation, 6.619 tok/s Prefill, TTFT 679 ms (27 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. ROG STRIX X570-F GAMING - 283,5 tok/s Generation, 3.276 tok/s Prefill, TTFT 3.524 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUS H7607BA - 146,9 tok/s Generation, 1.835 tok/s Prefill, TTFT 5.732 ms (4 Laufe)ASUS H7607BAASUSTeK COMPUTER INC. GX10 - 136,6 tok/s Generation, 2.246 tok/s Prefill, TTFT 5.100 ms (3 Laufe)ASUSTeK COMPUTER INC....ASUSTeK COMPUTER INC. PRIME A520M-K - 2.263,7 tok/s Generation, 8.103 tok/s Prefill, TTFT 10.659 ms (3 Laufe) | DIESER LAUF★ ASUSTeK COMPUTER INC....
ASUSTeK COMPUTER INC. ROG STRIX B550-A GAMING 3.484,3 tok/s★ ASUSTeK COMPUTER INC. PRIME A520M-K 2.263,7 tok/s this runASUSTeK COMPUTER INC. Pro WS WRX90E-SAGE SE 290,3 tok/sASUSTeK COMPUTER INC. ROG STRIX X570-F GAMING 283,5 tok/sASUS H7607BA 146,9 tok/sASUSTeK COMPUTER INC. GX10 136,6 tok/s

ENGby engine

4.4723.3542.2361.1180,06436.81412.98619.157Prefill (tok/s)Generation (tok/s)tensorfold - 3.484,3 tok/s Generation, 11.782 tok/s Prefill, TTFT 3.950 ms (13 Laufe)tensorfoldunbekannt - 859,8 tok/s Generation, 16.001 tok/s Prefill, TTFT 839 ms (6 Laufe)unbekanntllama.cpp - 286,2 tok/s Generation, 3.799 tok/s Prefill, TTFT 2.471 ms (37 Laufe)llama.cppvLLM - 2.263,7 tok/s Generation, 13.170 tok/s Prefill, TTFT 3.120 ms (13 Laufe) | DIESER LAUF★ vLLM
tensorfold 3.484,3 tok/s★ vLLM 2.263,7 tok/s this rununbekannt 859,8 tok/sllama.cpp 286,2 tok/s

DRVby driver

4.5023.3772.2511.1260,06663.7276.7899.850Prefill (tok/s)Generation (tok/s)unbekannt - 3.484,3 tok/s Generation, 8.270 tok/s Prefill, TTFT 2.622 ms (66 Laufe)unbekanntNVIDIA 590.48.01 / CUDA 13.1 - 136,6 tok/s Generation, 2.246 tok/s Prefill, TTFT 5.100 ms (3 Laufe)NVIDIA 590.48.01 / CU...
unbekannt 3.484,3 tok/sNVIDIA 590.48.01 / CUDA 13.1 136,6 tok/s
💰 Economics

Economics of this run

Operating cost, TCO and comparison with the next-best runs of the same model at identical concurrency (5× 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 25 W
⚡ TDP 450 W
☁️ External LLM (API)
Electricity0.30 EUR/kWh
Avg power (incl. idle)450 W estimated (TDP)GPU 450 W full load
Avg cost / hourEUR 0.14
Electricity / 1M tokensEUR 0.033
Token / kWh9.08M
Acquisition (system)EUR 1,270 partial priceGPU EUR 999 · CPU EUR 121 · PSU EUR 150
Electricity (2 years)–
TCO (2 years)EUR 3,635
Output tokens (2 years)71.58B
☁️ 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 (25 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

NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090 TiNVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090NVIDIA-Nemotron-3.5-Lightning-30B-A3BNVIDIA GeForce RTX 3090
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

Marcel Sommer

@marcelsommer