The card is a 300 W-class workstation part, not a silent HTPC dongle. Plan the box like a small workstation.
Official R9700 figures (AMD product page)
| Item | AMD-published |
|---|---|
| TBP | 300 W |
| PSU recommendation | 750 W |
| Power connector | 12V-2x6 |
| Slot | Double slot, active cooling |
| Memory | 32 GB GDDR6, 640 GB/s peak |
| Bus | PCIe 5.0 x16 |
| ECC | Linux only |
Street prices change; we do not list them. Partner coolers change acoustics. Look at the actual AIB board.
PSU
750 W is AMD’s minimum recommendation for the card in a generic system, not “headroom for a 5090-class CPU and six drives.” Use a quality unit with a native 12V-2x6 / 12VHPWR cable if that is what the card expects. Cheap pigtails are how people learn about transient spikes.
Cooling and noise
Many AI PRO cards are compact blower designs that exhaust out the bracket. Good for a cramped workstation; loud at the desk when memory clocks sit at the top P-state. Leave the adjacent slot empty. Do not bury the intake in cables. Watch junction and memory temps with rocm-smi while generating, not at the GNOME wallpaper.
Open-air test benches lie about case airflow. A small cube case will teach you about recirculation.
Case and risers
A short dual-slot PRO card fits places a triple-slot 7900 XTX does not. A PCIe 4.0 x16 riser will usually work. Decode-bound chat may not care; fat prefill might. Prefer a direct x16 slot if you are chasing prefill.
VRAM versus model size (approximate)
Weights + KV cache + batch share 32 GB. Verify per file and context. These are planning brackets, not promises:
| Situation | 32 GB often means |
|---|---|
| Qwen3.6-35B-A3B UD-Q5_K_XL, 262K, f16 KV | Our one-card daily driver — see stacks |
| Same at Q8 / even longer cache | Tight; measure. Floor is still Q4. |
| Qwen 3.8 Flash UD-Q6_K_XL | Publisher file ~169 GB — expect RAM to hold most of it on 32 GB |
| Dense 70B | Aggressive quant, short context, or another GPU |
| MoE with few active experts | File size can surprise you — read the card and the calculator |
A 16 GB RX 9070 XT is the better buy if you live in smaller Qwen. The R9700 is the 32 GB bracket. Never drop below Q4 to “make it fit.”
Host RAM: our test rig is 256 GB DDR5. 64 GB is a reasonable home minimum next to one 32 GB GPU. Flash at UD-Q6_K_XL is a ~169 GB publisher file — that box needs a lot of RAM. 32 GB host gets ugly fast.
Check price
No live prices. As an Amazon Associate I earn from qualifying purchases.
Check price · no live or stale prices
AMD Radeon AI PRO R9700
Check price · no live or stale prices
High-capacity DDR5 kit
Search only — not a specific SKU. Our test rig is 256 GB DDR5; 64 GB is a reasonable one-card minimum.
Check price · no live or stale prices
Fast NVMe SSD
House rule — load GGUFs from NVMe, not a spinning disk. Search only, not a SKU.
Check price · no live or stale prices
1000W-class PSU
AMD’s R9700 minimum recommendation is 750 W. A 1000W-class unit is search guidance for headroom, not a SKU.
As an Amazon Associate I earn from qualifying purchases. Some Newegg links use Rakuten Advertising and some B&H links use B&H’s affiliate program when those IDs are configured. Micro Center and AMD Store links are plain “also available at” links — no tracking. Buttons say “Check price” and never show a live or stale price.Full disclosure.
What not to optimize first
Do not buy a second card for “double tok/s” before you have one GPU proven. Multi-GPU at home pays an RCCL/topology tax. See the weekly honesty list.