Adam

LLM benchmark database for AMD AI PRO, Radeon, and Ryzen AI Max GPUs

[Adam @dad__vibes]

R9700 + RX 9070 XT, llama.cpp layer split. MoE gen tok/s at 128K, CPU expert offload → none:

Gemma 4: 79.06 → 143.43
Laguna: 57.22 → 89.74
GLM: 45.52 → 74.80

Dense 27B at 32K, already fit: 30.27 → 31.57

Extra VRAM helps MoE, not dense.

via kr4ckhe4d: https://t.co/6Kanpo74Nh

[Adam @dad__vibes]

One R9700 32GB running Qwen3.8-27B MXFP4 + DFlash2 on SlyBase's vLLM fork (ROCm 10.1), full 262k context:

decode: 135.8 tok/s
prefill: 3,132 (1.5k) / 2,635 (47k) tok/s
8 users: 461 tok/s aggregate

Their earlier run: a 210 W cap cost ~6.5% decode.

https://t.co/72WaHYWZqu

[Adam @dad__vibes]

Day 4 of telling X that https://t.co/xamU6WRKf0 holds real LLM benchmarks for the R9700 and other AMD Pro cards.

[Adam @dad__vibes]

Radeon AI PRO R9700 32GB vs RTX A5000 24GB · Ollama, ROCm 7.1.1, Ubuntu 24.04

tok/s (max context)

• deepseek-r1:14b: 53.52 (80K) / 53.15 (48K)
• deepseek-r1:32b: 26.29 (36K) / 25.77 (12K)
• gpt-oss:20b MoE: 102.40 (128K) / 119.46 (128K)

Dense runs roughly even. The A5000 wins the MoE, but the extra 8GB gives the R9700 3× the context on 32B.

Numbers by @Hostkey
https://t.co/JlOeaW3G7q

[Adam @dad__vibes]

Dual Radeon AI PRO R9700 · Qwen3.8-27B

llama.cpp HIP default Q8_0: pp 590 / tg 19.2
gfx1201 rebuild + tensor split + FA: pp 1458 / tg 29.7
vLLM Radiance MXFP4: pp 3948 / tg 184.9

Stock decode sits ~20 tok/s. Radiance MXFP4 jumps it almost 10× to ~185.

Source: https://t.co/QdBzC5YUWM
https://t.co/I5m8msJ5wY

[Adam @dad__vibes]

Day 3: still reminding X that https://t.co/xamU6WRKf0 is a benchmark database for R9700, W7900 and other AMD Pro cards.

[Adam @dad__vibes]

Radeon AI PRO R9700 32GB · llama-bench, Qwen3 Q4_K_M, flash attn on, full offload

ROCm / Vulkan, tok/s

tg128
• 8B: 94.3 / 95.3
• 14B: 52.7 / 53.1
• 32B: 24.9 / 25.0

pp512
• 8B: 3891 / 3978
• 32B: 921 / 927

Backend barely matters on RDNA4. The 32GB is the real win: 32B fits, while the 16GB RX 9700 XT can't load it.

Numbers by @timmyitdotcom
https://t.co/E438bFcm39

[Adam @dad__vibes]

W7900 48GB, llama.cpp MTP on vs off

Qwen3.6-35B-A3B Q4_K_S, llama-server Vulkan, Win11, 8K ctx

code: 49.7 → 99.4 tok/s (2.00×)
reasoning: 49.5 → 92.3 (1.86×)
long gen: 49.9 → 83.7 (1.68×)

TTFT barely moved. Biggest gain on code.

Source: https://t.co/HCJhwf1Hyw

[Adam @dad__vibes]

Day 2 of telling X that https://t.co/xamU6WRKf0 is a benchmark database for the Radeon AI PRO R9700, R9600 and Radeon PRO W7000 cards.

[Adam @dad__vibes]

R9700 · llama.cpp Vulkan/RADV, flash attn, full offload

Qwen3-30B-A3B Q4_K_M
• pp512 ~3033 tok/s
• tg128 183 tok/s

Qwen3.5-35B-A3B UD-Q4_K_XL
• pp512 2713 tok/s
• tg 127.4 tok/s
(same model on RTX 5090 CUDA: 7026 / 194)

MoE stays interactive on one card. Decode ~1.5× behind a 5090.

https://t.co/klS9ndc7B1

[Adam @dad__vibes]

2x Radeon AI PRO R9700, Qwen3.6-27B, tok/s at 1 / 8 users:

vLLM FP8 TP=2: 15.9 / 156.2
Tuned vLLM + MTP: 62.8 / 320.2
llama.cpp Q4_K_M: 23.4 / 56.0

Solo chat, llama.cpp beats stock vLLM. Add users and vLLM pulls way ahead.

via @PugetSystems https://t.co/K96IKukpgl

[Adam @dad__vibes]

Day 1 of reminding X that https://t.co/xamU6WRKf0 is a benchmark database for R9700 and other AMD Pro cards.

[Adam @dad__vibes]

Need a second opinion on your code — who do you ask first, Claude or Codex?

[Adam @dad__vibes]

Sunday roll call: Team Claude or Team ChatGPT, and why?

[Adam @dad__vibes]

Saturday night with the kids finally asleep — firing up Claude or ChatGPT?

[Adam @dad__vibes]

Claude Code or Codex?

[Adam @dad__vibes]

I've been leaning Claude for weekend tinkering lately — you riding with Claude or Codex this weekend?

[Adam @dad__vibes]

Friday check-in: Opus 5.5 or GPT-6 Astra — which one actually earned its keep for you this week?

[Adam @dad__vibes]

Hump day night with the side project still open — Claude or GPT tonight?

[Adam @dad__vibes]

Which one do you hit the usage limit on first — Claude or Codex?