Hardware classes
What kind of local-AI machine is this?
None of these classes is universally better. Capacity, bandwidth, software, power, and setup pull in different directions.
Catalogue review recorded · individual claims have separate datesReferenceSourceMethodology
GPU workstation
Dedicated VRAM and vendor-specific software support. A runtime supporting NVIDIA does not imply equivalent support for AMD or Intel. 14 cards in the database.
- Capacity: VRAM only, unless the runtime offloads
- Speed: compare the same model, precision, context and batch
- Power: board TDP is not measured whole-system consumption
Apple Silicon
CPU and GPU share a memory pool with macOS. Runtime support can be official, community-maintained or experimental. 16 configurations are recorded separately by memory capacity.
- Capacity: installed memory is not the process allocation limit
- Bandwidth: inspect the exact chip and configuration
- Software: Metal or an Apple-specific runtime/plugin, not CUDA
CPU / RAM offload
Some runtimes place selected model layers or experts in system RAM. Execution then depends on CPU work, memory bandwidth and transfers as well as GPU performance.
- Capacity: can exceed VRAM
- Speed: requires measurements of the actual offload configuration
- Evidence: record split settings, context, concurrency and runtime
Software conditions: vLLM installation requirements distinguish native GPU support from its Apple Metal plugin. llama.cpp documents split modes; combined card capacity alone is not a tested fit or speed result.
First-party lab host
0 published measurements. Host records omit hostname, username, serial, MAC, and IP.
- MacBook Air (M5, 16 GB unified) — Apple M5 (4 performance + 6 efficiency cores); 16 GB RAM; 16 GB unified; macOS 26.6.2; metal.
For a whole-machine recipe rather than a parts list, use the PC build archetypes.