New Challenges on FHERMA for Accelerating FHE Using NVIDIA cuPQC

Fully Homomorphic Encryption (FHE) is a powerful Privacy-Enhancing Technology (PET) that enables computations on encrypted data without access to the secret key, with enormous promise to transform privacy technologies across healthcare, finance, machine learning and beyond. Unlike traditional encryption, which requires data to be decrypted before it can be processed, FHE allows computation to be performed directly on encrypted data — the underlying plaintext is never exposed, not even to the system running the computation. Yet despite significant progress, two challenges have kept it from widespread deployment: complexity and performance.

Fair Math and NVIDIA are working together to address the performance bottlenecks of Fully Homomorphic Encryption by developing high-performance, GPU-accelerated building blocks through open challenges on the FHERMA platform, leveraging the state-of-the-art NVIDIA cuPQC SDK.

The collaboration focuses on the core computational building blocks shared across modern FHE implementations. By GPU-accelerating these reusable components, the initiative aims to benefit the entire ecosystem, enabling existing libraries and future applications to build on a common foundation of high-performance GPU kernels. These kernels will form the basis of reusable open infrastructure for applications including private information retrieval, privacy-preserving machine learning, and other encrypted workloads where performance has historically been the barrier to deployment.

The Performance Problem

The primary obstacle to practical FHE is performance. Modern RLWE-based schemes rely on arithmetic over high-dimensional polynomials with coefficient moduli that span hundreds or even thousands of bits — well beyond the native integer width of modern processors. Every homomorphic operation therefore translates into millions of modular arithmetic operations, making encrypted computation significantly more expensive than its plaintext equivalent.

Bridging this performance gap requires both algorithmic innovation and highly optimized implementations. This is precisely where GPU acceleration and optimized cryptographic building blocks can make a significant difference.

Unlocking the Full Potential of FHE Using GPUs

GPUs are a natural target for this effort because they occupy a uniquely useful point between general-purpose CPUs and custom Application-Specific Integrated Circuit (ASIC) architectures. FHE workloads expose substantial parallelism through modular additions, multiplications and Number Theoretic Transforms (NTTs), which operate independently across thousands of polynomial coefficients simultaneously. CPUs cannot fully exploit this parallel structure, constrained by modest core counts and memory bandwidth for the dense polynomial arithmetic. ASICs can deliver excellent performance, but they require long design cycles, large chip area and fixed arithmetic choices that may render them incompatible with new, cutting-edge cryptographic schemes. GPUs sit between the two, offering massive parallel throughput, a flexible programming model and high memory bandwidth. That makes them particularly well-suited to rapidly implementing, benchmarking and deploying new FHE kernels while still running on widely available hardware.

NVIDIA cuPQC SDK

At the heart of this initiative is the NVIDIA cuPQC SDK, a GPU-native cryptographic math layer that gives developers high-performance primitives without the burden of building low-level infrastructure from scratch. The SDK now includes two new libraries: cuPQC-NTT, which provides GPU-native Number Theoretic Transform support, and cuPQC-BigInt, which brings multiprecision integer arithmetic to the GPU. Both are optimized for NVIDIA GPU architectures and designed to be used as flexible, reusable components, so developers can focus on FHE innovation rather than low-level kernel engineering. Together, they give the broader FHE community the GPU-native building blocks to tackle the most compute-intensive operations in any FHE workload.

FHERMA Challenges

FHERMA is a benchmark and challenge platform for fully homomorphic encryption that aims to create an open ecosystem of production-quality building blocks for FHE and other privacy-preserving technologies. By drawing on the state-of-the-art cuPQC SDK, the FHERMA competition-based development model will enable new GPU-accelerated components that can significantly improve FHE performance.

We are starting with two challenge tracks:

High-Degree Polynomial Arithmetic
CKKS Key Switching

The High-Degree Polynomial Arithmetic track focuses on multiplication of high-degree polynomials with large integer coefficients — one of the most computationally demanding operations in modern FHE schemes. Combining NVIDIA cuPQC's Big Integer library with FHERMA's competition-driven development model will lead to new GPU-accelerated polynomial arithmetic techniques that benefit the entire FHE ecosystem. Although the resulting components are intended for FHE, participating in this track does not require prior expertise in homomorphic encryption.

The CKKS Key Switching track targets one of the main performance bottlenecks of the CKKS encryption scheme, the de facto standard for privacy-preserving inference powered by FHE. Rather than a single challenge, this track consists of a series of challenges covering the core computational building blocks of the key switching algorithm, all built on the same Big Integer arithmetic backend.

The best implementations become part of an open ecosystem of FHE components that can be reused across different libraries and applications.

Get Started

Join the initiative on the FHERMA platform. Review the open challenges, build on the cuPQC SDK, and submit your implementation for evaluation to compete for prizes.