| --- |
| license: mit |
| task_categories: |
| - text-generation |
| tags: |
| - benchmark |
| - nvidia |
| - dgx-spark |
| - blackwell |
| - llm |
| - inference |
| pretty_name: DGX Spark LLM Benchmarks |
| size_categories: |
| - n<1K |
| --- |
| |
| # DGX Spark LLM Benchmarks |
|
|
| First comprehensive benchmark suite for NVIDIA DGX Spark (GB10 Blackwell). |
|
|
| ## Hardware |
| - **GPU:** NVIDIA GB10 Blackwell (1 PFLOP FP4) |
| - **Memory:** 128GB unified LPDDR5x (273 GB/s) |
| - **CPU:** 20-core ARM (10x Cortex-X925 + 10x Cortex-A725) |
| - **Storage:** 4TB NVMe |
| - **Framework:** Ollama 0.18.3 |
| - **CUDA:** 13.0 | Driver: 580.142 |
|
|
| ## Benchmark Results |
|
|
| ### Run 1 — General Inference (11 models) |
|
|
| | Model | Size | Prompt tok/s | Gen tok/s | Load Time | |
| |-------|------|-------------|-----------|-----------| |
| | Llama 3.1 8B | 4.9 GB | 574.79 | 42.86 | 4.73s | |
| | Gemma3 27B | 17 GB | 164.64 | 11.71 | 9.83s | |
| | Qwen2.5-Coder 32B | 19 GB | 288.96 | 10.36 | 15.13s | |
| | Qwen3 32B | 20 GB | 141.91 | 9.88 | 4.08s | |
| | CodeLlama 70B | 38 GB | 133.35 | 5.73 | 28.81s | |
| | Nemotron 70B | 42 GB | 87.97 | 4.77 | 27.25s | |
| | Llama 3.1 70B | 42 GB | 67.92 | 4.76 | 28.35s | |
| | DeepSeek-R1 70B | 42 GB | 24.18 | 4.68 | 47.02s | |
| | Llama 3.3 70B | 42 GB | 67.97 | 4.66 | 27.41s | |
| | Qwen 2.5 72B | 47 GB | 122.19 | 4.40 | 44.75s | |
| | Mistral Large 123B | 73 GB | 10.43 | 2.28 | 86.14s | |
|
|
| ### Run 2 — Coding Benchmark |
|
|
| | Model | Gen tok/s | Tokens Generated | |
| |-------|-----------|-----------------| |
| | Llama 3.1 8B | 42.51 | 369 | |
| | Qwen2.5-Coder 32B | 10.33 | 552 | |
| | Qwen3 32B | 9.38 | 8,180 | |
| | CodeLlama 70B | 5.69 | 752 | |
| | Gemma3 27B | 11.61 | 329 | |
| | DeepSeek-R1 70B | 4.51 | 10,710 | |
| | Llama 3.3 70B | 4.67 | 507 | |
|
|
| ### Run 3 — Context Scaling |
|
|
| | Model | Short Prompt tok/s | Long Prompt tok/s | Scale | Gen tok/s | |
| |-------|-------------------|-------------------|-------|-----------| |
| | Llama 3.1 8B | 574 | 2,090 | 3.6x | 41.82 | |
| | Gemma3 27B | 164 | 616 | 3.8x | 11.70 | |
| | Qwen3 32B | 141 | 491 | 3.5x | 10.10 | |
| | Llama 3.1 70B | 67 | 212 | 3.2x | 4.69 | |
| | Qwen 2.5 72B | 122 | 225 | 1.8x | 4.33 | |
| | Nemotron 70B | 87 | 164 | 1.9x | 4.62 | |
|
|
| ### Run 4 — Vision |
|
|
| | Model | Size | Prompt tok/s | Gen tok/s | Load Time | |
| |-------|------|-------------|-----------|-----------| |
| | Llama3.2-Vision 90B | 54 GB | 6.02 | 3.47 | 16.86s | |
|
|
| ## Key Findings |
|
|
| 1. **27-32B is the sweet spot** — 10-12 tok/s, genuinely interactive |
| 2. **Prompt eval scales 3-4x** with longer prompts on unified memory |
| 3. **DeepSeek-R1 generates 10,710 tokens** of reasoning for one coding question |
| 4. **90B vision model runs** on a desktop at 3.47 tok/s |
| 5. **123B is the ceiling** — Mistral Large at 2.28 tok/s barely interactive |
| 6. **Generation speed is constant** regardless of prompt length |
|
|
| ## Author |
|
|
| **Gopi Trinadh Maddikunta** |
| - University of Houston · MS Engineering Data Science |
| - Research Assistant, Dr. Peizhu Qian |
| - GSoC 2025 Contributor (Scala Center) |
| - GitHub: GOPITRINADH3561 |
| - Website: gopitrinadh.site |
|
|
