Fully Homomorphic Encryption Benchmarks 2026: CKKS and BFV

CKKS vs BFV performance in 2026

The landscape of fully homomorphic encryption (FHE) in 2026 is no longer defined by theoretical possibility but by practical specialization. CKKS (Cheon-Kim-Kim-Song) and BFV (Brakerski-Fan-Vercauteren) have diverged into distinct roles. CKKS dominates privacy-preserving machine learning, while BFV remains the standard for exact integer arithmetic. This split is driven by how each scheme handles mathematical precision and computational overhead.

CKKS is designed for approximate arithmetic, making it ideal for neural networks and statistical models where slight numerical errors are acceptable. In 2026, performance gains from hardware and software co-design have significantly reduced the latency of CKKS operations. Specialized accelerators, as noted by researchers at NYU Tandon, are now optimizing the heavy linear algebra operations that CKKS relies on, bringing encrypted inference times closer to practical deployment thresholds.

BFV, by contrast, supports exact integer arithmetic. It is the go-to choice for applications requiring precise results, such as secure voting systems or exact financial calculations. While BFV has historically suffered from higher computational overhead, 2026 benchmarks show that optimized bootstrapping techniques and improved parameter selection have narrowed the performance gap. However, it remains less efficient than CKKS for large-scale, continuous data processing.

The choice between CKKS and BFV ultimately depends on the mathematical nature of your workload. If your application involves floating-point numbers and machine learning, CKKS is the clear winner. For integer-based, exact computation, BFV remains the robust choice. The 2026 toolkit landscape reflects this divergence, with libraries offering specialized optimizations for each scheme.

Latency and throughput comparison

Choosing between CKKS and BFV in 2026 is no longer a theoretical exercise; it is a hardware-aware engineering decision. While both schemes operate on the same fundamental lattice-based principles, their structural differences in noise management and mathematical support create distinct performance profiles. CKKS, designed for approximate arithmetic, handles floating-point operations with a different noise budget than BFV, which excels at exact integer arithmetic. This divergence means that benchmarks from 2024 are largely obsolete, as new hardware-software co-designs have dramatically altered the latency-to-throughput ratio for both schemes.

The core tradeoff lies in how each scheme manages computational depth. CKKS allows for deeper circuits by treating noise as a manageable error margin, making it ideal for machine learning inference and statistical analysis. BFV, by contrast, requires stricter noise management to preserve exact results, limiting its depth but offering higher precision for smaller datasets. Recent accelerators, such as those tested in the FHEIns benchmark, show that specialized ASICs can achieve up to 24.7× speedups over general-purpose CPU implementations, but these gains are not uniform across both schemes.

The table below compares key performance metrics for CKKS and BFV under typical 2026 benchmark conditions. These figures reflect real-world throughput on modern hardware, highlighting where each scheme shines. CKKS generally offers higher throughput for large-scale approximate computations, while BFV maintains lower latency for smaller, exact integer operations.

These numbers illustrate a critical point: there is no universal "fastest" scheme. CKKS’s higher throughput comes at the cost of increased latency for individual operations, while BFV’s lower latency is constrained by its tighter noise budget. For applications requiring deep neural network inference, CKKS’s ability to handle larger circuits without premature noise exhaustion makes it the default choice. Conversely, for financial calculations or cryptographic protocols requiring exact results, BFV’s precision and lower per-operation latency remain superior.

The 2026 landscape also emphasizes hardware co-design. Accelerators optimized for CKKS’s polynomial multiplication patterns often struggle with BFV’s specific noise reduction steps, and vice versa. This specialization means that selecting the right scheme is inextricably linked to the underlying hardware architecture. As benchmarking tools become more sophisticated, developers must align their algorithmic choices with both the mathematical needs of their application and the hardware capabilities of their deployment environment.

Toolkit optimization trends

The 2026 landscape for Fully Homomorphic Encryption (FHE) has shifted from theoretical viability to engineering maturity. Libraries like HEaaN and OpenFHE have reduced computational overhead through aggressive SIMD optimizations and hardware acceleration, making FHE increasingly viable for onchain use. This section compares how CKKS and BFV implementations have evolved to meet distinct mathematical needs.

HEaaN, a high-performance CKKS library, focuses on numerical precision for machine learning workloads. Its 2026 optimizations leverage SIMD instructions to process multiple data elements in parallel, significantly reducing the latency of polynomial multiplication. This approach is ideal for applications requiring approximate arithmetic, such as privacy-preserving inference on encrypted datasets.

BFV, implemented robustly in OpenFHE, remains the standard for integer arithmetic and homomorphic operations on discrete values. OpenFHE’s recent updates introduce hardware-aware parameter selection, allowing developers to tune ciphertext sizes and noise budgets more efficiently. This co-design between software libraries and underlying hardware architectures has resulted in 2-5x speedups in bootstrapping operations compared to 2023 baselines.

The convergence of software and hardware optimization is evident in specialized accelerators. Projects like FHEIns demonstrate up to 24.7× speedups on database benchmarks by tailoring ASIC designs to FHE-specific operations. These gains are critical for onchain applications, where gas costs and execution time directly impact feasibility.

The choice between CKKS and BFV now depends less on raw performance and more on the specific mathematical requirements of the application. CKKS offers superior throughput for numerical data, while BFV provides a simpler, more predictable environment for integer-based logic. As hardware acceleration continues to mature, both schemes are becoming practical for real-world deployment.

Choosing for onchain compute

Selecting between CKKS and BFV for onchain workloads requires matching the mathematical operation to the constraint. CKKS handles approximate arithmetic, making it the standard for machine learning inference where small precision losses are acceptable. BFV uses exact integer arithmetic, which is necessary for secure voting or discrete aggregation where rounding errors are unacceptable.

In 2026, hardware and software co-design has narrowed the performance gap between these schemes. Newer libraries leverage SIMD parallelism to process multiple ciphertexts simultaneously, reducing the latency penalty that once made FHE impractical for blockchain nodes. The decision now hinges on whether your use case prioritizes numerical flexibility or strict data integrity.

For machine learning tasks, CKKS allows you to run inference on encrypted weights and inputs. The scheme’s ability to handle approximate arithmetic means you can perform matrix multiplications without the massive overhead of exact integer conversion. This makes it the only viable option for privacy-preserving AI models on constrained chains.

If your goal is secure voting or multi-party computation, BFV is the correct choice. It guarantees that no rounding occurs during addition or multiplication, preserving the exact outcome. While BFV ciphertexts are generally smaller, the scheme lacks the native support for the continuous operations required by neural networks.

Frequently asked: what to check next