Geerling ai-benchmarks: AMD Radeon AI Pro R9700 (in Intel 265K PC), llama.cpp Vulkan, DeepSeek-R1-14b (Q4 distill), 53.72 tok/s eval rate, 330.8W peak
53.72 tok/s eval rate
Fields
id
benchmarks/geerling-r9700-deepseekr1-14b-vulkan-llamacpp-decodeMetric
decode_tok_s
Value
53.72
Unit
tok/s eval rate
Model
deepseek-ai/DeepSeek-R1-Distill-Qwen-14B (Q4 GGUF)
Methodology
THIRD-PARTY MEASURED (Jeff Geerling, geerlingguy/ai-benchmarks issue #35). Host: Intel Core Ultra 265K custom PC. llama.cpp with Vulkan backend (footnote 1 in the Findings table marks the R9700 rows as llama.cpp/Vulkan, not Ollama). DeepSeek-R1-14b Q4 GGUF. Eval rate 53.72 tok/s; peak system power 330.8W. Per-request decode, batch 1.
Measured by
third_party
Reproducible
yes
Energy basis
wall power at the plug measured alongside the run; 330.8W peak reported in the table
Mean power (W)
330.8
Power scope
node
notes
The same R9700 card attached to a Pi CM5 collapses to 7.98 tok/s at 279.7W — the Pi's PCIe Gen3 x4 link plus weak host CPU cripples llama.cpp/Vulkan far worse than the RTX cards' Ollama path (4090 on Pi retains 90% of x86-host speed). Host-quality sensitivity is the story on AMD-Vulkan, not the card's raw capability.
Notes
The same R9700 card attached to a Pi CM5 collapses to 7.98 tok/s at 279.7W — the Pi's PCIe Gen3 x4 link plus weak host CPU cripples llama.cpp/Vulkan far worse than the RTX cards' Ollama path (4090 on Pi retains 90% of x86-host speed). Host-quality sensitivity is the story on AMD-Vulkan, not the card's raw capability.