FHE lowered to tensor ops
CKKS and BGV primitives, NTT, key switching, rescaling, bootstrapping, expressed as batched tensor kernels that map cleanly onto modern accelerators.
HEAL · Homomorphic Encryption Abstraction Layer
CUDA for FHE. One stack. Every accelerator.
HEAL is Lattica's hardware abstraction layer for Fully Homomorphic Encryption. It lowers FHE primitives to tensor operations and dispatches them to whatever accelerator is fastest, GPU today, TPU, FPGA, and FHE specific ASICs tomorrow, without rewriting the workload.
THE LAYER
What HEAL is, and why it matters
Classical FHE libraries hard code one cryptographic scheme into one CPU implementation. Every new hardware target means rewriting the world. HEAL breaks that lock-in.
Hardware agnostic by design
A single backend interface. Add a new accelerator by implementing a small set of tensor primitives, the rest of the stack, compiler, runtime, SDK, comes along for free.
Built for parallel silicon
Massively parallel batching across ciphertexts, pipelined bootstrapping, and memory aware scheduling, designed to saturate the hardware, not waste it.
THE UNLOCK
FHE is linear algebra
Accelerators love linear algebra
HEAL is the bridge
WHERE HEAL SITS
Between the cryptography and the silicon.
HEAL sits between FHE schemes and the hardware that runs them. It is the enabler that lets FHE workloads target the best available accelerator without rewriting the stack.
L1 · Applications
FHE Applications
Encrypted DB
AI Inference
Smart Contracts
Custom pipelines
L2 · Software
FHE Software Stack
Schemes
BGV
CKKS
TFHE
GSW
Building blocks
Polynomial arithmetic
Key switching
Bootstrapping
L3 · Abstraction Layer
HEAL
Universal Interface for FHE Operations
Hardware-agnostic IR
Instruction scheduling
Runtime & dispatch
Vendor extensions
L4 · Hardware
Acceleration Hardware
GPU
TPU
FPGA
FHE ASIC
HARDWARE TARGETS
Write once. Run on the fastest silicon.
HEAL's backend interface is small, well defined, and built to be extended. New accelerators plug in without disturbing the rest of the stack.
GPU
Production todayFirst class backend. Validated on data center NVIDIA GPUs, sub-second encrypted inference on real workloads.
TPU
On the roadmapTPUs are purpose-built for tensor workloads, making them a perfect target for HEAL's tensor-based abstraction. Backend in active development.
FPGA
Partner programReconfigurable fabrics are ideal for the modular arithmetic at the heart of FHE. Open to silicon partners.
FHE-specific ASICs
Partner programHEAL's tensor IR is the natural target for purpose-built FHE silicon. We co-design with chip teams.
THE TOOLKIT
Everything you need to build a backend.
An open IR specification, an open-source runtime, a reference CPU backend, a conformance suite, and a public benchmarking path - all in the open.
Documentation
HEAL Docs
IR specification, backend interface, architecture guides.
healdocs.lattica.aiSource
Open-source on GitHub
Runtime, reference CPU backend, examples. Read the code.
Lattica-ai/healTesting
Conformance & test tooling
Validate a new backend against the HEAL contract end-to-end.
View on GitHubBenchmarking
FHE Benchmarking Standard
Lattica defined how remote-backend FHE is benchmarked for the community standard. A HEAL backend lets your silicon submit on the same footing.
fhe-benchmarking.github.ioTalk to the HEAL team
Building accelerator silicon, GPU, TPU, FPGA, or a purpose-built FHE chip? Tell us about your hardware and we’ll set up a technical deep dive with our compiler and runtime team.