What has to be rewritten?
Anything that assumes plaintext at the accelerator. RainDB, ZChat, and the SDK are the supported surfaces. Existing SQL or inference stacks that decrypt into GPU memory are not drop-in.
Integrate
The uncomfortable questions, answered.
Which parts of my pipeline still see plaintext?
The client encrypt/decrypt step, and key custody. Umbra compute does not. If you add GPUs in a mixed inference pack for unencrypted stages, those GPUs see plaintext. Draw that boundary in the design review.
Who holds the keys?
You. ZipLogic does not operate customer data and does not hold customer keys.
What model can I run, and who supplies it?
You supply the model and the data. ZipLogic supplies the card, runtime, and supported surfaces. We do not run a hosted model API on your holdings.
What happens when the card fails?
It is a host-attached PCIe card. Replace it. N4/N8 nodes have multiple cards; a single-card failure is a capacity event, not a key event. Keys never lived on the failed card as a recoverable plaintext store.
How do updates work in an air-gapped facility?
Manually, by your team. There is no vendor-initiated remote update path.
What is the actual latency on my workload class?
It depends on the workload. Card throughput is 40M encrypted logic operations per second, measured. End-to-end latency is not a single number. Request the methodology pack and characterise on one card before you size a node.