Datasets:

Scherm-AI
/

onprem-llm-benchmark

	---
	license: cc-by-4.0
	tags:
	- benchmark
	- llm-inference
	- on-premise
	- vllm
	- gpu
	pretty_name: Scherm On-Premise LLM Inference Benchmark
	---

	# Scherm On-Premise LLM Inference Benchmark — v0.7

	Throughput, latency and resource benchmarks for on-premise LLM inference across a wide range
	of GPUs — datacenter (B200, L40S), Blackwell workstation (RTX PRO 6000), consumer (RTX 4090/4070),
	unified-memory (DGX Spark / GB10) and legacy/AMD — covering the full stack of a production
	assistant: chat serving, embeddings, reranking, vision, OCR and speech-to-text.

	Engine: vLLM (mostly the NGC `26.03` image, engine v0.17.1; Blackwell needs it for sm_120/121
	kernels). Each measurement is taken on a fully dedicated GPU (`--exclusive`), inference only
	(server already warm — no model load/download counted).

	→ See [`REPRODUCING.md`](REPRODUCING.md) to reproduce every experiment on your own hardware.
	Raw per-experiment CSVs live under [`raw/<machine>/`](raw/); consolidated tables under [`data/`](data/).

	### 🚀 Run it on your own Slurm cluster (NVIDIA or AMD)

	Self-contained jobs in [`scripts/portable-slurm/`](scripts/portable-slurm/) recreate everything from
	scratch — they download the runners straight from this repo, no setup of ours required. Four flavors
	(`nvidia_container` · `nvidia_venv` · `amd_container` · `amd_venv`); AMD/ROCm covers Frontier/ORNL.

	```bash
	# preview every job that would be submitted (nothing runs):
	DRY_RUN=1 SBATCH_SCRIPT=amd_container.sbatch PARTITION=batch bash scripts/portable-slurm/run_all.sh
	# then submit all benchmarks (all models × all tasks) with one chosen sbatch:
	SBATCH_SCRIPT=amd_container.sbatch PARTITION=batch ACCOUNT=<projid> \
	SCRATCH=/lustre/orion/<projid>/scratch/$USER bash scripts/portable-slurm/run_all.sh
	```

	See [`scripts/portable-slurm/README.md`](scripts/portable-slurm/README.md) for which flavor to pick
	and what to adapt to your cluster.

	---

	## 🏆 Ranking de GPU por modelo (pico de throughput, tok/s)

	> AWQ = 4-bit quantizado; BF16 = precisão cheia. A maioria das GPUs roda AWQ (cabe mais); a
	> RTX PRO 6000 (96GB) roda ambos — a comparação AWQ×BF16 aparece onde há as duas.

	### Qwen2.5-7B (chat)
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA B200 \| AWQ \| 17470 \| 1024 \|
	\| 2 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 7762 \| 128 \|
	\| 3 \| NVIDIA RTX PRO 6000 Blackwell \| BF16 \| 6496 \| 128 \|
	\| 4 \| NVIDIA L40S \| AWQ \| 3504 \| 64 \|
	\| 5 \| NVIDIA L40S (Grace ARM) \| AWQ \| 3121 \| 128 \|
	\| 6 \| NVIDIA RTX 4090 \| AWQ \| 2710 \| 128 \|
	\| 7 \| NVIDIA GB10 (DGX Spark) \| AWQ \| 1291 \| 64 \|
	\| 8 \| NVIDIA RTX 4070 \| AWQ \| 1085 \| 64 \|

	### Qwen2.5-14B (chat)
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA B200 \| AWQ \| 8758 \| 512 \|
	\| 2 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 3573 \| 128 \|
	\| 3 \| NVIDIA RTX PRO 6000 Blackwell \| BF16 \| 1746 \| 128 \|
	\| 4 \| NVIDIA L40S \| AWQ \| 1660 \| 64 \|
	\| 5 \| NVIDIA L40S (Grace ARM) \| AWQ \| 1469 \| 128 \|
	\| 6 \| NVIDIA RTX 4090 \| AWQ \| 984 \| 32 \|
	\| 7 \| NVIDIA GB10 (DGX Spark) \| AWQ \| 422 \| 32 \|

	### Qwen2.5-32B (chat)
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA B200 \| AWQ \| 5089 \| 512 \|
	\| 2 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 1735 \| 128 \|
	\| 3 \| NVIDIA RTX PRO 6000 Blackwell \| BF16 \| 1156 \| 128 \|
	\| 4 \| NVIDIA L40S \| AWQ \| 852 \| 64 \|
	\| 5 \| NVIDIA L40S (Grace ARM) \| AWQ \| 516 \| 32 \|
	\| 6 \| NVIDIA RTX 4090 \| AWQ \| 392 \| 16 \|
	\| 7 \| NVIDIA GB10 (DGX Spark) \| AWQ \| 297 \| 64 \|

	### Qwen2.5-72B (≥80GB)
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA B200 \| AWQ \| 2672 \| 512 \|
	\| 2 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 860 \| 128 \|

	### Qwen2.5-Coder-7B
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA B200 \| AWQ \| 13832 \| 256 \|
	\| 2 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 7763 \| 128 \|
	\| 3 \| NVIDIA RTX PRO 6000 Blackwell \| BF16 \| 3708 \| 128 \|
	\| 4 \| NVIDIA L40S \| AWQ \| 3491 \| 64 \|
	\| 5 \| NVIDIA L40S (Grace ARM) \| AWQ \| 3172 \| 128 \|
	\| 6 \| NVIDIA RTX 4090 \| AWQ \| 2023 \| 128 \|
	\| 7 \| NVIDIA GB10 (DGX Spark) \| AWQ \| 1172 \| 64 \|
	\| 8 \| NVIDIA RTX 4070 \| AWQ \| 857 \| 64 \|

	### Qwen2.5-Coder-14B
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA B200 \| AWQ \| 7040 \| 512 \|
	\| 2 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 3572 \| 128 \|
	\| 3 \| NVIDIA L40S \| AWQ \| 1646 \| 64 \|
	\| 4 \| NVIDIA L40S (Grace ARM) \| AWQ \| 1502 \| 128 \|
	\| 5 \| NVIDIA RTX 4090 \| AWQ \| 990 \| 64 \|
	\| 6 \| NVIDIA GB10 (DGX Spark) \| AWQ \| 554 \| 64 \|

	### Qwen2.5-Coder-32B
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA B200 \| AWQ \| 4577 \| 512 \|
	\| 2 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 1735 \| 128 \|
	\| 3 \| NVIDIA L40S \| AWQ \| 867 \| 64 \|
	\| 4 \| NVIDIA GB10 (DGX Spark) \| AWQ \| 299 \| 64 \|

	### Llama-3.3-70B (AWQ)
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 391 \| 128 \|

	### Mixtral-8x7B (AWQ)
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA RTX PRO 6000 Blackwell \| AWQ \| 3229 \| 128 \|

	### Llama-Guard-3-8B (guardrail)
	\| # \| GPU \| quant \| tok/s \| @conc \|
	\|---\|-----\|-------\|-------\|-------\|
	\| 1 \| NVIDIA RTX PRO 6000 Blackwell \| BF16 \| 5258 \| 128 \|
	\| 2 \| NVIDIA L40S (Grace ARM) \| BF16 \| 2096 \| 128 \|

	---

	## 🧩 Stack do assistente além do chat

	### Embeddings — `BAAI/bge-m3` (embeddings/s, pico)
	\| # \| GPU \| emb/s \|
	\|---\|-----\|-------\|
	\| 1 \| NVIDIA RTX PRO 6000 Blackwell \| 1706 \|
	\| 2 \| NVIDIA L40S (Grace ARM) \| 729 \|
	\| 3 \| NVIDIA RTX 4090 \| 727 \|
	\| 4 \| NVIDIA L40S \| 685 \|
	\| 5 \| NVIDIA RTX 4070 \| 291 \|
	\| 6 \| NVIDIA RTX 2080 Ti \| 61 \|
	\| 7 \| NVIDIA GB10 (DGX Spark) \| 59 \|
	\| 8 \| NVIDIA GTX 1080 Ti \| 27 \|
	\| 9 \| AMD RX 7900 XT \| 11 \|

	### Reranking — `BAAI/bge-reranker-v2-m3` (pairs/s, pico)
	\| # \| GPU \| pairs/s \|
	\|---\|-----\|---------\|
	\| 1 \| NVIDIA RTX PRO 6000 Blackwell \| 6407 \|
	\| 2 \| NVIDIA RTX 4090 \| 4000 \|
	\| 3 \| NVIDIA L40S (Grace ARM) \| 3681 \|
	\| 4 \| NVIDIA RTX 4070 \| 2132 \|

	### Speech-to-text — `whisper-large-v3` (faster-whisper, audio-seconds/s, pico)
	\| # \| GPU \| audio_s/s \|
	\|---\|-----\|-----------\|
	\| 1 \| NVIDIA RTX PRO 6000 Blackwell \| 449 \|
	\| 2 \| NVIDIA RTX 4090 \| 343 \|
	\| 3 \| NVIDIA RTX 4070 \| 182 \|

	> STT roda só em GPUs x86 — o L40S aqui é Grace (ARM/aarch64), onde CTranslate2/faster-whisper
	> não compila. RTF medido < 0,02 (transcreve muito mais rápido que tempo real).

	### Vision — `Qwen2.5-VL-7B-Instruct` (tok/s, pico)
	\| # \| GPU \| tok/s \|
	\|---\|-----\|-------\|
	\| 1 \| NVIDIA RTX 4090 \| 2264 \|
	\| 2 \| NVIDIA L40S (Grace ARM) \| 2067 \|
	\| 3 \| NVIDIA RTX PRO 6000 Blackwell \| 1404 \|
	\| 4 \| NVIDIA L40S \| 1171 \|

	---

	## 📄 OCR de documentos — latência por página (doc 1 página)

	Extração de documentos com modelos de visão (vLLM). Página A4 jurídica típica.

	\| Modelo \| GPU \| ms/página \| tok/s \| nota \|
	\|--------\|-----\|-----------\|-------\|------\|
	\| DeepSeek-OCR (3B) \| RTX PRO 6000 \| 320 \| 512 \| 🥇 OCR-direto, ~11× mais rápido que NuMarkdown \|
	\| DeepSeek-OCR-2 (3B) \| RTX PRO 6000 \| 394 \| 413 \| quase igual \|
	\| Qwen2.5-VL-7B \| RTX 4090 \| 1148 \| 56 \| VLM genérico \|
	\| NuMarkdown-8B-Thinking \| RTX PRO 6000 \| 3554 \| 101 \| thinking model (gera raciocínio/página) \|
	\| NuMarkdown-8B-Thinking \| RTX 4090 \| 5520 \| 57 \| — \|

	> DeepSeek-OCR vs NuMarkdown: na RTX PRO 6000, o DeepSeek-OCR (3B) é ~11× mais rápido
	> (320 vs 3554 ms/página). O NuMarkdown é um thinking model de 8B (Qwen2.5-VL por baixo) que gera
	> centenas de tokens de raciocínio por página antes do markdown; o DeepSeek-OCR é 3B otimizado pra
	> OCR direto. Para extração em volume, DeepSeek-OCR é o caminho. *(Acurácia CER/WER pendente de
	> ground-truth — o texto extraído de cada modelo está salvo para comparação.)*

	### NuMarkdown-8B em todas as GPUs (baseline cross-GPU, doc 1 página)
	Dados consolidados em [`data/ocr_pages_ALL.csv`](data/ocr_pages_ALL.csv); brutos em [`raw/`](raw/).

	\| GPU \| VRAM \| ms/pág \| tok/s \| doc 50 págs (total) \|
	\|-----\|------\|--------\|-------\|---------------------\|
	\| B200 \| 180GB \| 2133 \| 152 \| 167s (~3min) 🥇 \|
	\| RTX PRO 6000 \| 96GB \| 3554 \| 101 \| 555s (~9min) \|
	\| RTX 4090 \| 24GB \| 5520 \| 57 \| 842s (~14min) 🥉 melhor custo \|
	\| L40S \| 48GB \| 6424 \| 49 \| 518s (~9min) \|
	\| DGX Spark (GB10) \| 128GB \| 25617 \| 13 \| 1981s (~33min) 🐌 memory-bound \|

	> Sizing pro OCR: banda de memória > VRAM total. B200 (HBM3e) faz OCR a 152 tok/s — 12× a
	> Spark. A RTX 4090 (consumer) é ~4,5× mais rápida que a DGX Spark no decode do thinking-model.
	> OCR pesado nunca deve rodar na Spark. GPUs Pascal/Turing e a RTX 4070 (12GB, OOM) não rodam
	> OCR-8B.

	---

	## 💰 Custo-benefício — tok/s por US$1.000 (Qwen2.5-32B AWQ, faixa de produção)

	\| GPU \| preço aprox. \| tok/s (pico) \| tok/s por US$1k \|
	\|-----\|-------------\|--------------\|-----------------\|
	\| RTX PRO 6000 Blackwell \| $9.000 \| 1735 \| 193 🥇 \|
	\| RTX 4090 \| $3.549 \| 392 \| 110 \|
	\| L40S \| $8.300 \| 852 \| 103 \|
	\| DGX Spark (GB10) \| $4.699 \| 297 \| 63 \|
	\| B200 (1 GPU) \| ~$40.000 \| 5089 \| 127 \|

	> A RTX PRO 6000 lidera o custo-benefício single-GPU na faixa de produção (32B-AWQ): cabe o modelo
	> com folga em 96GB e entrega o melhor tok/s por dólar. O B200 é imbatível em throughput absoluto,
	> mas a um custo por GPU ~4,5× maior.

	---

	## 🗺️ Matriz de cobertura — o que rodou em cada GPU

	\| GPU \| serving \| embed \| rerank \| stt \| vision \| ocr \|
	\|-----\|:-------:\|:-----:\|:------:\|:---:\|:------:\|:---:\|
	\| B200 \| ✅ \| — \| — \| — \| — \| ✅ \|
	\| RTX PRO 6000 Blackwell \| ✅ \| ✅ \| ✅ \| ✅ \| ✅ \| ✅ \|
	\| L40S \| ✅ \| ✅ \| ✅ \| ⛔ARM \| ✅ \| ✅ \|
	\| RTX 4090 \| ✅ \| ✅ \| ✅ \| ✅ \| ✅ \| ✅ \|
	\| RTX 4070 \| ✅ \| ✅ \| ✅ \| ✅ \| ⛔12GB \| ⛔12GB \|
	\| GB10 (DGX Spark) \| ✅ \| ✅ \| — \| — \| — \| ✅ \|
	\| RTX 2080 Ti / 1080 Ti / P100 \| ollama \| ✅ \| ⛔ \| ⛔ \| ⛔ \| ⛔ \|
	\| AMD RX 7900 XT \| ollama \| ✅ \| ⛔ROCm \| ⛔ROCm \| ⛔ROCm \| ⛔ROCm \|

	Tetos físicos (não são lacunas, são limites de hardware):
	- AMD RX 7900 XT — vLLM-ROCm só faz serving de texto; pooling (rerank), vision e CTranslate2-STT não têm caminho ROCm.
	- L40S (Grace) — aarch64; faster-whisper/CTranslate2 não compila → sem STT.
	- GPUs pré-2020 (P100, 1080Ti, 2080Ti) — sem vLLM moderno; só ollama (serving) + embeddings.
	- RTX 4070 (12GB) — não cabe modelos de visão/OCR de 7-8B (OOM).

	---

	## 📁 Estrutura do repositório

	```
	data/ *_ALL.csv consolidados (schema único, 1 linha por ponto medido)
	raw/ CSV bruto por experimento, por máquina (gppd/ b200/ spark/ rtx-6000/)
	scripts/ runners + jobs por plataforma (runners/ gppd-slurm/ b200-apptainer/ rtx-docker/ spark-ollama/)
	REPRODUCING.md passo-a-passo para reproduzir em cada hardware
	```

	## ⚠️ Nota sobre engines

	A maioria dos resultados usa a imagem NGC vLLM (`26.03`, engine v0.17.1) para uniformidade e suporte
	Blackwell. Alguns runs antigos usaram vLLM v0.11.0; DeepSeek-OCR usa a imagem oficial `vllm/vllm-openai`
	(a NGC não traz `torchaudio`, dependência da inspeção da arquitetura). Para comparação justa entre
	GPUs, prefira linhas com o mesmo engine.