FHE Hardware
Advances in specialized hardware (ASICs)
While algorithms boosts FHE speed 8x per year, the hardware developments compound this effect. There is growing interest in designing FHE-specific hardware.
ASIC stands for Application-Specific Integrated Circuit. They are specialized hardwares, or you can call custom-designed chips, that are built to perform well-defined set of operations. The mainstream CPUs are designed to be reasonably performant across many types of computations. If you know what kind of math operations you will do beforehand (and in what sequence you will do), you can tailor your chip to be very performant on those.
The GPUs themselves are ASICs to CPUs. They are designed to be very performant on matrix multiplications (matmuls). That means, GPUs are also advantageous for FHE computations which involve lots of matmuls. But FHE, in addition to matmuls, also relies on the following operations: - Modular Arithmetic - Chinese Remainder Theory (CRT) based operations - Polynamial Arithmetic - Number-Theoric Transforms (NTT)
Therefore, there are opportunities for FHE ASICs that can perform better than generic GPUs, by optimizing above operations.
An example ASIC design (HE-Booster on GPU) is below. Notice that it includes modules for CRT, NTT, and some other HE-specific operations.
FHE ciphertexts are 40-1000x larger than the plaintext, meaning that ASICs need to have lots of memory. Additionally, FHE is embarrassingly parallel (a feature of lattice-based math) that creates other opportunies for special hardware.
Hardware acceleration
![[A High-Level Technical Overview of Fully Homomorphic Encryption (archived)#Hardware acceleration]]
FHE Hardware Startup Ecosystem
A new wave of companies is building FHE-specific acceleration:
Niobium Microsystems - Raised $5.5M in 2024 for FHE accelerator chips - Focus: Custom ASICs optimized for lattice-based cryptography - Potential: 1000-10,000× speedup for FHE operations
Fabric Cryptography - Building custom silicon for cryptographic workloads - FHE-optimized processors with specialized arithmetic units - Targeting cloud deployment and edge computing
Agita Labs - Hardware acceleration for privacy-preserving computation - Focus on making FHE practical for real-time applications
Duality Technologies - Software-hardware co-design approach - Real deployments in healthcare (patient privacy) and finance (fraud detection) - Notable: CTO founded the Palisade Library, chief cryptographer developed BGV scheme
From Fully Homomorphic Encryption to Silicon - What is Microsoft's HEAX? (archived)
https://x.com/i/grok/share/zUMo4KqWBszye3FbM4Qv8ozjW
https://www.jeremykun.com/2024/05/04/fhe-overview/#hardware-acceleration
Chips to Compute With Encrypted Data Are Coming (archived) - FHE Hardware-accelerator startups: - https://niobiummicrosystems.com/ - https://www.fabriccryptography.com/ - https://agitalabs.com/ - https://dualitytech.com/ - preserving patient privacy in healthcare - financial firms, check for fraud - The CTO of Duality is also the founder of the Palisade Library while their chief cryptographer is the developer of a leading FHE scheme called BGV.
FHE Market Landscape
- https://x.com/Dod_2206/status/1943312631227650410
- TFHE-rs (github)
- https://docs.zama.ai/tfhe-rs
- FHE Optimization Opportunities — - Increasing interest in hardware acceleration for FHE. See [[FHE Hardware]]
- Fully Homomorphic Encryption and the Dawn of A Truly Private Internet — Alongside with algorithm improvements, hardware improvements are also happening, compounding the effect. More on that...
- Fully Homomorphic Encryption and the Dawn of a Truly Private Internet (v0) — [[FHE Hardware]]
- Fully Homomorphic Encryption and the Dawn of a Truly Private Internet (v1) — [[FHE Hardware]]
- The Coming Wave of Encrypted Computation (v2) — [[FHE Hardware]]
- Chips to Compute With Encrypted Data Are Coming (archived)
- FHE Market Landscape
- From Fully Homomorphic Encryption to Silicon - What is Microsoft's HEAX? (archived)
- HE-Booster An Efficient Polynomial Arithmetic Acceleration on GPUs for Fully Homomorphic Encryption (archived)
- UDC2018 Sep.14 Jung Hee Cheon Security & Smart Contract Ⅰ (archived)
- Chinese Remainder Theory (CRT)
- Number-Theoric Transforms (NTT)
- Niobium Microsystems
- Fabric Cryptography
- Agita Labs
- Duality Technologies
- https://x.com/i/grok/share/zUMo4KqWBszye3FbM4Qv8ozjW
- https://www.jeremykun.com/2024/05/04/fhe-overview/#hardware-acceleration
- https://www.biometricupdate.com/202405/niobium-raises-5-5m-to-develop-fully-homomorphic-encryption-accelerator-chip
- Apheris AI
- https://cornami.com/
- https://x.com/Dod_2206/status/1943312631227650410
- github
- https://docs.zama.ai/tfhe-rs