
Product
A card that uses locked data.
Run AI and database work without opening the files. The server never sees them in the clear.
Thesis
Software proved you can compute on a lock. Hardware makes it usable in a rack.
Umbra is ZipLogic’s flagship Encrypted Processing Unit: a hardware acceleration platform specifically engineered for homomorphic execution.
Powered by our proprietary, noise-free ZX-Engine, Umbra replaces deep, noisy computational chains with structured lookup tables, driving per-operation processing costs down toward ordinary arithmetic.
Umbra computes on encrypted data.
40,000,000 encrypted operations per second1. Available as a card, or as a configured node.
Architecture
The Umbra architecture: ZipLogic’s flagship Encrypted Processing Unit
Umbra is a dedicated hardware acceleration platform built to execute frontier artificial intelligence and relational analytics natively over ciphertext. By moving homomorphic computation off general-purpose host GPUs and onto specialized, non-decrypting silicon arrays, Umbra enforces an absolute cryptographic boundary: server memory holds static noise, while the customer retains exclusive custody of the secret key.
Why purpose-built
Engineered specifically for homomorphic execution
General-purpose GPUs are designed for high-throughput floating-point arithmetic on unencrypted data. When forced to run traditional Fully Homomorphic Encryption, GPUs collapse under immense memory bottlenecks, continuous cache flushes, and mathematical noise accumulation.
Umbra bypasses these physical limits by introducing a purpose-built execution topology designed from the circuit level up for noise-free homomorphic logic.
Data path
- 01Customer key domain
Encrypt at the edge. Secret keys never leave custody.
- 02Host · ciphertext only
Server memory holds static noise: not plaintext.
- 03Umbra EPU · ZX-Engine
Noise-free homomorphic logic via structured lookup tables.
- 04Encrypted result
Outputs stay ciphertext until deliberate decryption by the owner.
Product data
U100 card · N8 node

| Platform | Umbra Encrypted Processing Unit |
|---|---|
| Engine | ZX-Engine · noise-free LUT topology |
| Card throughput | 40,000,000 encrypted logic ops/s1 |
| N8 node | 320,000,000 ops/s estimated · ~1.4 kW wall1 |
| Host interface | PCIe ×8 Gen3 |
| Form factor | Full-height, full-length, single-slot |
| Attachment | Host-attached: no on-card network interface |
| Runtime | Umbra runtime · Umbra SDK |

Deploy
Card path or finished system
Integrate U100 into your own servers, or take N4 / N8 when ZipLogic owns thermals, burn-in, and acceptance. Same runtime either way.
See where it sits in a rack you already own.
Methodology
1 Encrypted logic operations per second, measured on production silicon. Node figures estimated from measured single-card throughput.