[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f31oet3pn9674h":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":36,"bodyZh":67,"bodyEn":68,"sourceName":69,"sourceUrl":70,"updatedAt":71},"rtx-5090-desktop-benchmarks-2026-10-11","NEWS","2026-10-11T12:00:00.000Z","RTX 5090 桌面实测复盘：35 条新记录平均 270 tok\u002Fs，15.1GB VRAM","RTX 5090 desktop benchmark recap: 35 new records avg 270 tok\u002Fs, 15.1GB VRAM","10-10 批量 RTX 5090 (32GB) 35 条 L2 记录平均 270.04 tok\u002Fs，平均占用 15.11GB vram。RTX 5090 是当前桌面卡 200+ tok\u002Fs 俱乐部领头羊，比 RTX 4090 (174 tok\u002Fs) 高 55%。","10-10 batch RTX 5090 (32GB) 35 L2 records avg 270.04 tok\u002Fs, avg 15.11GB vram used. RTX 5090 is current desktop 200+ tok\u002Fs club leader, 55% above RTX 4090 (174 tok\u002Fs).",[12,15,18,21,24,27,30,33],{"id":13,"name":14},5,"Qwen3.8-27B",{"id":16,"name":17},6,"Qwen3.6-27B",{"id":19,"name":20},48,"Qwen3.5-4B",{"id":22,"name":23},50,"Qwen3.5-27B",{"id":25,"name":26},51,"Qwen3.5-0.8B",{"id":28,"name":29},53,"Qwen3-8B",{"id":31,"name":32},55,"Qwen3-1.7B",{"id":34,"name":35},65,"gemma-4-12B-it",[37,40,43,46,49,52,55,58,61,64],{"id":38,"name":39},1,"NVIDIA RTX 4090",{"id":41,"name":42},2,"NVIDIA RTX 3090",{"id":44,"name":45},4,"Apple M3 Max",{"id":47,"name":48},11,"AMD Radeon RX 7900 XTX",{"id":50,"name":51},13,"NVIDIA RTX 5090",{"id":53,"name":54},14,"NVIDIA RTX 5080",{"id":56,"name":57},15,"NVIDIA RTX 3060 12GB",{"id":59,"name":60},18,"NVIDIA RTX 5070",{"id":62,"name":63},26,"Apple M4 Max",{"id":65,"name":66},94,"Intel Arc B580","# RTX 5090 桌面实测复盘：35 条新记录平均 270 tok\u002Fs，15.1GB VRAM\n\n10-10 批量 35 条 L2 实测基于 NVIDIA RTX 5090 (32GB GDDR7)——首次给桌面卡有大规模样本。\n\n## 整体复盘\n\n| 指标 | 数值 |\n|---|---|\n| 记录数 | 35 |\n| 平均 decode tok\u002Fs | **270.04** |\n| 平均 VRAM 占用 | **15.11 GB**（46% of 32GB）|\n| 模型范围 | 0.8B → 27B |\n| 最高单条 | **384.44 tok\u002Fs**（Qwen3.5-0.8B + Q8_0）|\n| 最低单条 | 122.20 tok\u002Fs（Qwen3.5-27B + Q2_K）|\n\n**关键观察**：平均 VRAM 15.11GB = 32GB 卡的 46%——**说明当前 27B 模型在 5090 上仍有大量 VRAM 富余**。这给跑 70B+ 量化版留了空间。\n\n## 横向对比桌面卡\n\n10-10 批量中 6 个桌面 \u002F 笔记本 GPU 的平均表现：\n\n| 硬件 | VRAM | 记录数 | 平均 tok\u002Fs | 相对 5090 |\n|---|---|---|---|---|\n| **RTX 5090** | 32GB | 35 | **270.04** | 基准 |\n| Apple M4 Max | 128GB (unified) | 47 | 196.99 | -27% |\n| RTX 3090 | 24GB | 18 | 185.06 | -31% |\n| RTX 4090 | 24GB | 23 | 174.22 | -35% |\n| RTX 5080 | 16GB | 16 | 165.90 | -39% |\n| RTX 5070 | 12GB | 17 | 147.10 | -46% |\n| RTX 3060 12GB | 12GB | 16 | 137.98 | -49% |\n| Apple M3 Max | 36GB (unified) | 34 | 152.53 | -44% |\n| AMD RX 7900 XTX | 24GB | 21 | 147.03 | -46% |\n| Intel Arc B580 | 12GB | 16 | 95.72 | -65% |\n\n**RTX 5090 桌面卡王座确认**——比第二名 Apple M4 Max 高 37%，比 RTX 4090 高 55%。\n\n## 5 个最戏剧性的 RTX 5090 单条\n\n| 模型 | 量化 | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3.5-0.8B | Q8_0 | **384.44** | 4.03 GB |\n| Qwen3.5-27B | IQ1_S | 260.56 | 25.91 GB |\n| Qwen3-1.7B | Q5_K_M | 353.77 | 5.42 GB |\n| Qwen3.5-4B | Q3_K_M | 356.10 | 7.81 GB |\n| Qwen3.5-4B | Q4_K_S | 353.67 | 8.65 GB |\n\n**Qwen3.5-0.8B + Q8_0 + RTX 5090 = 384 tok\u002Fs**——这是什么概念：每秒能读完一本 6000 字短篇小说，单卡桌面，无人值守。\n\n## 装机建议\n\n如果预算 ¥14000+，RTX 5090 是当前桌面推理的\"甜区\"：\n\n| 预算 | 方案 | 27B tok\u002Fs | 4B tok\u002Fs |\n|---|---|---|---|\n| ¥14000 | RTX 5090 + 64GB DDR5 + NVMe 4.0 | **175+** | **356** |\n| ¥14000 | RTX 4090 + 64GB DDR5 + NVMe 4.0 | 91 | 230 |\n| ¥5500 | RTX 5070 + 64GB DDR5 + NVMe 4.0 | 79 | 200 |\n\n**多花 ¥8500 升级到 5090**：27B 提速 92%，4B 提速 78%。\n\n## 证据分层\n\n- 270.04 平均：DB 35 条，hardware_id=13\n- 384.44 最高：DB record id=1380\n- 横向对比 6 卡：DB `benchmark_record` 10-10 批量按 hardware_id 分组\n- VRAM 占用 15.11GB：DB 实际 vram_gb 字段平均\n- \"5090 比 4090 高 55%\"：270.04 \u002F 174.22 = 1.55\n\n## 数据来源\n\n- 35 条 RTX 5090 记录：DB `benchmark_record` 表，hardware_id=13\n- 横向对比：DB benchmark_record 10-10 批量按 hardware_id 分组\n- 最高单条：DB record id=1380（model_id=51, framework_id=3, quantization=Q8_0）","# RTX 5090 desktop benchmark recap: 35 new records avg 270 tok\u002Fs, 15.1GB VRAM\n\n10-10 batch has 35 L2 records on NVIDIA RTX 5090 (32GB GDDR7) — the first large-sample data for the desktop card.\n\n## Overall recap\n\n| Metric | Value |\n|---|---|\n| Records | 35 |\n| Avg decode tok\u002Fs | **270.04** |\n| Avg VRAM used | **15.11 GB** (46% of 32GB) |\n| Model range | 0.8B → 27B |\n| Best single record | **384.44 tok\u002Fs** (Qwen3.5-0.8B + Q8_0) |\n| Worst single record | 122.20 tok\u002Fs (Qwen3.5-27B + Q2_K) |\n\n**Key observation**: average VRAM 15.11GB = 46% of 32GB — **plenty of headroom for 70B+ quantized runs**.\n\n## Side-by-side desktop comparison\n\n6 desktop \u002F laptop GPUs from 10-10 batch:\n\n| Hardware | VRAM | Records | Avg tok\u002Fs | vs 5090 |\n|---|---|---|---|---|\n| **RTX 5090** | 32GB | 35 | **270.04** | baseline |\n| Apple M4 Max | 128GB (unified) | 47 | 196.99 | -27% |\n| RTX 3090 | 24GB | 18 | 185.06 | -31% |\n| RTX 4090 | 24GB | 23 | 174.22 | -35% |\n| RTX 5080 | 16GB | 16 | 165.90 | -39% |\n| RTX 5070 | 12GB | 17 | 147.10 | -46% |\n| RTX 3060 12GB | 12GB | 16 | 137.98 | -49% |\n| Apple M3 Max | 36GB (unified) | 34 | 152.53 | -44% |\n| AMD RX 7900 XTX | 24GB | 21 | 147.03 | -46% |\n| Intel Arc B580 | 12GB | 16 | 95.72 | -65% |\n\n**RTX 5090 confirmed as desktop king** — 37% above 2nd place (Apple M4 Max), 55% above RTX 4090.\n\n## 5 most dramatic single RTX 5090 records\n\n| Model | Quant | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3.5-0.8B | Q8_0 | **384.44** | 4.03 GB |\n| Qwen3.5-27B | IQ1_S | 260.56 | 25.91 GB |\n| Qwen3-1.7B | Q5_K_M | 353.77 | 5.42 GB |\n| Qwen3.5-4B | Q3_K_M | 356.10 | 7.81 GB |\n| Qwen3.5-4B | Q4_K_S | 353.67 | 8.65 GB |\n\n**Qwen3.5-0.8B + Q8_0 + RTX 5090 = 384 tok\u002Fs** — read a 6000-character short story per second, single card, no oversight.\n\n## Build recommendation\n\nIf budget is ¥14000+, RTX 5090 is the current \"sweet spot\" for desktop inference:\n\n| Budget | Setup | 27B tok\u002Fs | 4B tok\u002Fs |\n|---|---|---|---|\n| ¥14000 | RTX 5090 + 64GB DDR5 + NVMe 4.0 | **175+** | **356** |\n| ¥14000 | RTX 4090 + 64GB DDR5 + NVMe 4.0 | 91 | 230 |\n| ¥5500 | RTX 5070 + 64GB DDR5 + NVMe 4.0 | 79 | 200 |\n\n**Extra ¥8500 for 5090**: 27B +92%, 4B +78%.\n\n## Evidence\n\n- 270.04 average: DB 35 records, hardware_id=13\n- 384.44 best: DB record id=1380\n- Side-by-side 6 cards: DB benchmark_record 10-10 batch grouped by hardware_id\n- VRAM 15.11GB: DB actual vram_gb field average\n- \"5090 is 55% above 4090\": 270.04 \u002F 174.22 = 1.55\n\n## Sources\n\n- 35 RTX 5090 records: DB `benchmark_record` table, hardware_id=13\n- Side-by-side: DB benchmark_record 10-10 batch grouped by hardware_id\n- Best single: DB record id=1380 (model_id=51, framework_id=3, quantization=Q8_0)","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Frtx-5090-desktop-benchmarks-2026-10-11","2026-10-12T12:12:45.228Z"]