AMD Radeon AI PRO R9700 32GB workstation first measurements: 33 records avg 210 tok/s
10-13 batch 33 R9700 L2 records avg 210.14 tok/s (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.
Source: araoai
AMD Radeon AI PRO R9700 32GB workstation first measurements: 33 records avg 210 tok/s
10-13 batch has 33 AMD Radeon AI PRO R9700 (32GB GDDR6) L2 records — AMD workstation card first time in DB.
Overall recap
| Metric | Value |
|---|---|
| Records | 33 |
| Avg decode tok/s | 210.14 |
| Peak decode tok/s | 394.48 |
| Avg VRAM used | 9.46 GB (30% of 32GB) |
| Model range | 0.5B → 31B |
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.
R9700 vs same-price / same-tier N-cards
| Hardware | VRAM | Records | Avg tok/s | Price tier |
|---|---|---|---|---|
| RTX 5090 | 32GB | 35 | 270.04 | ¥14000 |
| AMD R9700 | 32GB | 33 | 210.14 | ¥9000 (est.) |
| AMD RX 7900 XTX | 24GB | 21 | 147.03 | ¥6500 |
| RTX 4080 | 16GB | 23 | 202.72 | ¥8500 |
| RTX 4090 | 24GB | 23 | 174.22 | ¥12000 |
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.
R9700 5 most dramatic single records
| Model | Quant | tok/s | VRAM |
|---|---|---|---|
| Qwen3.5-0.8B | IQ2_XS | 394.48 | 2.91 GB |
| Qwen3.5-0.8B | Q2_K | 380+ | 3.0 GB |
| Qwen2.5-0.5B-Instruct | Q5_K_M | 384.75 | 3.08 GB |
| Qwen3-1.7B | Q4_K_M | 280+ | 5.0 GB |
| Qwen3.5-27B | Q4_K_M | 95 | 22 GB |
Qwen3.5-0.8B IQ2_XS + R9700 + llama.cpp = 394.48 tok/s — one of the all-time fastest single records in DB.
Key insight: AMD inference ecosystem catching up
| Dimension | AMD R9700 | RTX 5090 |
|---|---|---|
| Inference speed | 210 tok/s | 270 tok/s |
| ROCm compatibility | ✅ 6.4+ full support | N/A |
| llama.cpp support | ✅ Vulkan + ROCm | ✅ CUDA |
| Strata support | ✅ from 0.5.3 | ✅ |
| vLLM support | ✅ ROCm | ✅ CUDA |
| Price | ¥9000 (est.) | ¥14000 |
AMD in 2026 has inference ecosystem that can compete with N-cards — 3 years ago the 7900 XTX couldn't make this comparison.
Build recommendation
| Budget | Setup | 27B tok/s | Recommendation |
|---|---|---|---|
| ¥9000 | R9700 + 64GB DDR5 + NVMe 4.0 | 95 | ⭐⭐⭐⭐ value |
| ¥12000 | RTX 4090 + 64GB DDR5 + NVMe 4.0 | 91 | ⭐⭐⭐ classic |
| ¥14000 | RTX 5090 + 64GB DDR5 + NVMe 4.0 | 175 | ⭐⭐⭐⭐⭐ top |
R9700 is the only ¥9000-tier card that runs 27B at full speed + has 32GB VRAM headroom.
Evidence
- 210.14 average: DB 33 records, hardware_id=79
- 394.48 peak: DB record id=1450
- 9.46GB VRAM average: DB actual vram_gb field
- ¥9000 price: AMD official 2025-09 pricing + domestic discounts
- ROCm 6.4+: AMD official docs
Sources
- 33 R9700 records: DB
benchmark_recordtable, hardware_id=79 - Peak single: DB record id=1450 (Qwen3.5-0.8B + IQ2_XS + llama.cpp)