Builds / RX 9070 XT 16 GB + Radeon AI PRO R9700 32 GB
RX 9070 XT + Radeon AI PRO R9700 (48 GB, ROCm 7.2): Qwen3.8 27B 71.6 tok/s with MTP at 128K context
- GPU(s)
- RX 9070 XT 16 GB + Radeon AI PRO R9700 32 GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 48
- CPU
- AMD Ryzen 7 9800X3D
Amazon · eBay (used) · Newegg - Motherboard
- not stated
- RAM
- 32 GB DDR5-6000
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Both cards PCIe 5.0 x8; CachyOS, ROCm 7.2; llama.cpp build 11345 (a868c3e3c), some rows b11434; layer split across the two cards
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 27B | IQ4_XS-v3, q8_0 KV | llama.cpp b11345 (ROCm), layer split, no drafter | 31.49 | not stated | context: 128K ctx, empty context📏 700-token generation probe, median of 3 |
| Qwen3.8 27B | IQ4_XS-v3 + MTP head, q8_0 KV | llama.cpp (ROCm), layer split, MTP | 71.59 | 1,014 @ 128K depth-token prompt | ⚡ spec-decode (MTP)context: 128K ctx, empty context📏 700-token generation probe, median of 3; 85.1% draft acceptance |
| Gemma 4 31B | Q6_K, f16 KV | llama.cpp b11434 (ROCm), layer split, no drafter | 19.3 | not stated | context: 128K ctx📏 700-token generation probe, median of 3 |
| Gemma 4 31B | Q6_K + MTP drafter, f16 KV | llama.cpp b11434 (ROCm), layer split, MTP | 52.2 | 279 @ 128K depth-token prompt | ⚡ spec-decode (MTP)context: 128K ctx📏 700-token generation probe, median of 3; 86.4% draft acceptance |
| Gemma 4 26B-A4B | Q4 (UD-Q4_K_M per the repo KLD table) + MTP | llama.cpp (ROCm), layer split, MTP | 143.43 | 1,203 @ 128K depth-token prompt | ⚡ spec-decode (MTP)context: 128K ctx📏 700-token generation probe, median of 3 |
| Qwen3.6 35B-A3B | UD-Q6_K | llama.cpp b11434 (ROCm), layer split, no speculative decoding | 71.99 | 3,170 @ 128K depth-token prompt | context: 128K ctx📏 700-token generation probe, median of 3 |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Owner-measured benchmarks in the kr4ckhe4d/local-llm-benchmarks repo (Oct 2026). Prefill figures are at 128K context depth, so they are much lower than empty-context prefill. Board and PSU not stated.
https://github.com/kr4ckhe4d/local-llm-benchmarks
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.