NL-2610.010Ongoing · Planned[nl.EB]
CUDA, Vulkan or the processor, measured properly
The Nook team
Opened 5 October 2026 · 1 min read
Our first look at audio.cpp (NL-2610.001) found its CUDA build more than twice as fast as its Vulkan build, from one run of each. Nook installs different builds of its engines depending on the computer: CUDA for NVIDIA cards, Vulkan for other cards, and processor builds when there is no usable card. We will time a fixed job on each build of llama.cpp, whisper.cpp, stable-diffusion.cpp and audio.cpp, many times over, and report medians, spread and peak memory, so the choice of build rests on more than one run.
- Stage
- Planned
- Comments
- RTX 4060 8 GB for CUDA and Vulkan, and the processor; AMD and Intel results welcome from readers.
- Subjects
- Engines and builds (nl.EB); Performance (nl.PF)
- Cite as
NL-2610.010· usenook.ai/labs/cuda-vulkan-cpu-across-engines.html
Full text
The question
How much does the build of an engine matter, engine by engine? A single run said the CUDA build of audio.cpp speaks more than twice as fast as its Vulkan build on our card. Is the gap as large for a chat model, for speech to text, for pictures? And how much slower is the processor alone, for someone without a usable graphics card?
The plan
- One fixed job per engine: a chat answer of fixed length (llama.cpp), a ten-minute recording (whisper.cpp), a 1024 × 1024 picture (stable-diffusion.cpp), and the voice job from NL-2610.001 (audio.cpp).
- Each job on each build: CUDA and Vulkan on the RTX 4060, and the processor build.
- Ten runs each after a warm-up run; we report the median, the spread, and peak memory.
- The download size of each build, since that is part of the choice.
So far
Planned. The single runs from NL-2610.001 are all we have. We only have an NVIDIA card, so Vulkan on AMD and Intel cards needs readers: if you run one, we would like your numbers.