[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2i2g06a8prjon":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":27,"bodyZh":43,"bodyEn":44,"sourceName":45,"sourceUrl":46,"updatedAt":47},"amd-r9700-workstation-debut-2026-10-14","NEWS","2026-10-14T10:00:00.000Z","AMD Radeon AI PRO R9700 32GB 工作站卡首次实测：33 条记录平均 210 tok\u002Fs","AMD Radeon AI PRO R9700 32GB workstation first measurements: 33 records avg 210 tok\u002Fs","10-13 批量 33 条 R9700 L2 实测平均 210.14 tok\u002Fs（峰值 394.48）——AMD 工作站卡首次进 DB。比 RTX 5090 慢 22%，但平均 VRAM 占用只 9.46GB，说明 R9700 的 32GB 富余很多。","10-13 batch 33 R9700 L2 records avg 210.14 tok\u002Fs (peak 394.48) — AMD workstation card first time in DB. 22% slower than RTX 5090, but avg VRAM usage only 9.46GB — 32GB has plenty of headroom.",[12,15,18,21,24],{"id":13,"name":14},48,"Qwen3.5-4B",{"id":16,"name":17},50,"Qwen3.5-27B",{"id":19,"name":20},51,"Qwen3.5-0.8B",{"id":22,"name":23},55,"Qwen3-1.7B",{"id":25,"name":26},59,"Qwen2.5-0.5B-Instruct",[28,31,34,37,40],{"id":29,"name":30},1,"NVIDIA RTX 4090",{"id":32,"name":33},11,"AMD Radeon RX 7900 XTX",{"id":35,"name":36},13,"NVIDIA RTX 5090",{"id":38,"name":39},41,"NVIDIA RTX 4080",{"id":41,"name":42},79,"AMD Radeon AI PRO R9700","# AMD Radeon AI PRO R9700 32GB 工作站卡首次实测：33 条记录平均 210 tok\u002Fs\n\n10-13 批量入库 33 条 AMD Radeon AI PRO R9700 (32GB GDDR6) L2 实测——**AMD 工作站卡首次进入本站 DB**。\n\n## 整体复盘\n\n| 指标 | 数值 |\n|---|---|\n| 记录数 | 33 |\n| 平均 decode tok\u002Fs | **210.14** |\n| 峰值 decode tok\u002Fs | **394.48** |\n| 平均 VRAM 占用 | **9.46 GB**（30% of 32GB）|\n| 模型范围 | 0.5B → 31B |\n\n**关键观察**：平均 VRAM 9.46GB 是 32GB 卡的 30%——R9700 在当前 27B 及以下模型上有 **70% 的 VRAM 富余**。给未来 70B 量化版留了巨大空间。\n\n## R9700 vs 同价位 \u002F 同档 N 卡\n\n| 硬件 | VRAM | 记录数 | 平均 tok\u002Fs | 价格档 |\n|---|---|---|---|---|\n| RTX 5090 | 32GB | 35 | 270.04 | ¥14000 |\n| **AMD R9700** | **32GB** | **33** | **210.14** | **¥9000**（推测）|\n| AMD RX 7900 XTX | 24GB | 21 | 147.03 | ¥6500 |\n| RTX 4080 | 16GB | 23 | 202.72 | ¥8500 |\n| RTX 4090 | 24GB | 23 | 174.22 | ¥12000 |\n\n**R9700 比 5090 慢 22%，比 4080 快 4%，比 7900 XTX 快 43%**——AMD 工作站卡在推理性能上已经追平 N 卡中高端。\n\n## R9700 5 个最戏剧性的单条\n\n| 模型 | 量化 | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3.5-0.8B | IQ2_XS | **394.48** | 2.91 GB |\n| Qwen3.5-0.8B | Q2_K | 380+ | 3.0 GB |\n| Qwen2.5-0.5B-Instruct | Q5_K_M | 384.75 | 3.08 GB |\n| Qwen3-1.7B | Q4_K_M | 280+ | 5.0 GB |\n| Qwen3.5-27B | Q4_K_M | 95 | 22 GB |\n\n**Qwen3.5-0.8B IQ2_XS + R9700 + llama.cpp = 394.48 tok\u002Fs**——本站 DB 历史最快单条之一。\n\n## 关键洞察：AMD 推理生态在追平\n\n| 维度 | AMD R9700 | RTX 5090 |\n|---|---|---|\n| 推理速度 | 210 tok\u002Fs | 270 tok\u002Fs |\n| ROCm 兼容性 | ✅ 6.4+ 全支持 | N\u002FA |\n| llama.cpp 支持 | ✅ Vulkan + ROCm | ✅ CUDA |\n| Strata 支持 | ✅ 0.5.3 起 | ✅ |\n| vLLM 支持 | ✅ ROCm | ✅ CUDA |\n| 价格 | ¥9000（推测）| ¥14000 |\n\n**AMD 在 2026 年的推理生态已经能和 N 卡叫板**——3 年前的 7900 XTX 没人敢这么比。\n\n## 装机建议\n\n| 预算 | 方案 | 27B tok\u002Fs | 推荐度 |\n|---|---|---|---|\n| ¥9000 | R9700 + 64GB DDR5 + NVMe 4.0 | 95 | ⭐⭐⭐⭐ 性价比 |\n| ¥12000 | RTX 4090 + 64GB DDR5 + NVMe 4.0 | 91 | ⭐⭐⭐ 经典 |\n| ¥14000 | RTX 5090 + 64GB DDR5 + NVMe 4.0 | 175 | ⭐⭐⭐⭐⭐ 顶级 |\n\n**R9700 是 ¥9000 价位唯一能跑 27B 全速 + 32GB VRAM 富余的卡**。\n\n## 证据分层\n\n- 210.14 平均：DB 33 条，hardware_id=79\n- 394.48 峰值：DB record id=1450\n- 9.46GB VRAM 平均：DB 实际 vram_gb 字段\n- 价格 ¥9000：基于 AMD 官方 2025-09 定价 + 国内电商折扣\n- ROCm 6.4+：AMD 官方文档\n\n## 数据来源\n\n- 33 条 R9700 记录：DB `benchmark_record` 表，hardware_id=79\n- 峰值单条：DB record id=1450（Qwen3.5-0.8B + IQ2_XS + llama.cpp）","# AMD Radeon AI PRO R9700 32GB workstation first measurements: 33 records avg 210 tok\u002Fs\n\n10-13 batch has 33 AMD Radeon AI PRO R9700 (32GB GDDR6) L2 records — **AMD workstation card first time in DB**.\n\n## Overall recap\n\n| Metric | Value |\n|---|---|\n| Records | 33 |\n| Avg decode tok\u002Fs | **210.14** |\n| Peak decode tok\u002Fs | **394.48** |\n| Avg VRAM used | **9.46 GB** (30% of 32GB) |\n| Model range | 0.5B → 31B |\n\n**Key observation**: average VRAM 9.46GB = 30% of 32GB — R9700 has **70% VRAM headroom** for 27B and below models. Plenty of room for future 70B quantized.\n\n## R9700 vs same-price \u002F same-tier N-cards\n\n| Hardware | VRAM | Records | Avg tok\u002Fs | Price tier |\n|---|---|---|---|---|\n| RTX 5090 | 32GB | 35 | 270.04 | ¥14000 |\n| **AMD R9700** | **32GB** | **33** | **210.14** | **¥9000** (est.) |\n| AMD RX 7900 XTX | 24GB | 21 | 147.03 | ¥6500 |\n| RTX 4080 | 16GB | 23 | 202.72 | ¥8500 |\n| RTX 4090 | 24GB | 23 | 174.22 | ¥12000 |\n\n**R9700 is 22% slower than 5090, 4% faster than 4080, 43% faster than 7900 XTX** — AMD workstation cards now rival N-card mid-to-high tier on inference.\n\n## R9700 5 most dramatic single records\n\n| Model | Quant | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3.5-0.8B | IQ2_XS | **394.48** | 2.91 GB |\n| Qwen3.5-0.8B | Q2_K | 380+ | 3.0 GB |\n| Qwen2.5-0.5B-Instruct | Q5_K_M | 384.75 | 3.08 GB |\n| Qwen3-1.7B | Q4_K_M | 280+ | 5.0 GB |\n| Qwen3.5-27B | Q4_K_M | 95 | 22 GB |\n\n**Qwen3.5-0.8B IQ2_XS + R9700 + llama.cpp = 394.48 tok\u002Fs** — one of the all-time fastest single records in DB.\n\n## Key insight: AMD inference ecosystem catching up\n\n| Dimension | AMD R9700 | RTX 5090 |\n|---|---|---|\n| Inference speed | 210 tok\u002Fs | 270 tok\u002Fs |\n| ROCm compatibility | ✅ 6.4+ full support | N\u002FA |\n| llama.cpp support | ✅ Vulkan + ROCm | ✅ CUDA |\n| Strata support | ✅ from 0.5.3 | ✅ |\n| vLLM support | ✅ ROCm | ✅ CUDA |\n| Price | ¥9000 (est.) | ¥14000 |\n\n**AMD in 2026 has inference ecosystem that can compete with N-cards** — 3 years ago the 7900 XTX couldn't make this comparison.\n\n## Build recommendation\n\n| Budget | Setup | 27B tok\u002Fs | Recommendation |\n|---|---|---|---|\n| ¥9000 | R9700 + 64GB DDR5 + NVMe 4.0 | 95 | ⭐⭐⭐⭐ value |\n| ¥12000 | RTX 4090 + 64GB DDR5 + NVMe 4.0 | 91 | ⭐⭐⭐ classic |\n| ¥14000 | RTX 5090 + 64GB DDR5 + NVMe 4.0 | 175 | ⭐⭐⭐⭐⭐ top |\n\n**R9700 is the only ¥9000-tier card that runs 27B at full speed + has 32GB VRAM headroom**.\n\n## Evidence\n\n- 210.14 average: DB 33 records, hardware_id=79\n- 394.48 peak: DB record id=1450\n- 9.46GB VRAM average: DB actual vram_gb field\n- ¥9000 price: AMD official 2025-09 pricing + domestic discounts\n- ROCm 6.4+: AMD official docs\n\n## Sources\n\n- 33 R9700 records: DB `benchmark_record` table, hardware_id=79\n- Peak single: DB record id=1450 (Qwen3.5-0.8B + IQ2_XS + llama.cpp)","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Famd-r9700-workstation-debut-2026-10-14","2026-10-15T13:12:56.417Z"]