Confidential computing & FHE
Computing on encrypted data without decryption
FPGA accelerates FHE primitives (NTT/CRT), CPU orchestrates circuits, GPU pre-processes matrices.
FPGA
Accelerated FHE primitives
Workloads
- ▸ BFV/CKKS NTT/INTT
- ▸ Key-switching
- ▸ Bootstrapping
Performance
×100 speedup vs CPU
CPU
Circuit orchestration
Workloads
- ▸ MLIR circuit compilation
- ▸ FHE context management
- ▸ OpenFHE / Concrete API
Performance
Circuits up to 1 bn gates
GPU
Preprocessing + encrypted ML
Workloads
- ▸ Plaintext batch encoding
- ▸ Encrypted NN inference
- ▸ Private distances
Performance
CNN inference on ciphertext
Multi-agent scenario
A hospital submits an encrypted scan: GPU pre-encodes, FPGA runs the accelerated FHE convolution without decrypting, CPU orchestrates the circuit and returns the encrypted result to the client.