The field guide / 02
A useful next step.
Work through a real problem, understand the limits, and leave with a better test.
Build a baseline you can trust.
Choose your software path, establish device visibility, and verify the actual response before chasing tokens per second.
Choose a coherent B70 software stack.
Pick a host environment and an application route that belong together, from GGUF inference to serving and creative work.
A symptom needs a version.
A dated troubleshooting log that keeps affected environments, upstream outcomes, and local verification separate.
Build a repeatable creative workflow.
Take images and video from a promising Intel XPU demo to a saved workflow, an inspectable output, and an honest timing record.
When does a second B70 help?
Separate memory capacity, independent jobs, and model splitting before you decide what two cards should do.
CUDA assumptions, meet Intel.
A practical decision guide for native XPU, library substitution, source translation, and the dependencies that stop a port.
A documentation review is different from a new GPU test. Each guide states its evidence date, environment, and remaining verification work.