Skip to main content
Get StartedBuild with zkTLSFHE

Confidential Finance Product

image

Confidential Finance Product

Primus Confidential Finance is a decentralized application for on-chain asset privacy, built with Fully Homomorphic Encryption (FHE), running on BNB Chain. Users can convert their stablecoin into the encrypted version, then deposit it into the confidential vault to earn yield, or send the encrypted asset to other wallet, without ever exposing balances, position size or transaction amount on-chain.

Users can access the platform for:

  • Shield / Unshield: Convert stablecoin into its encrypted version, or convert it back at any time. This is where the privacy-preserving stablecoin flow begins and ends.
  • Stake / Redeem: Deposit idle encrypted asset into the Confidential Vault to earn yield, with position size and strategy kept private end to end. Redemption is equally privacy-preserving at any time, as the encrypted asset is returned from the external vault.
  • Send: Confidential transfers of any encrypted stablecoin to any wallet on BNB Chain are also supported, without exposing the transfer amount or changed balances.

User Guide​

For a step-by-step walkthrough of the Confidential Finance product, including how to Shield, Stake, Send, and Redeem encrypted assets, see the User Guide.

The Performance​

The product is powered by an in-house FHE stack built end to end at Primus: a cryptographic scheme, a compiler, and a GPU engine.

Fully Homomorphic Encryption (FHE) allows computation to happen directly on encrypted data — balances, amounts, and logic never need to be decrypted, yet the results are still correct. In gate-based FHE, encrypted logic is evaluated as a boolean circuit, and each gate historically required its own bootstrapping, the refresh step that keeps ciphertext noise under control. A 64-bit operation compiles to thousands of gates, so the naive approach meant thousands of expensive bootstrappings per computation. Since bootstrapping is programmable, as it can evaluate a small encrypted function as part of the refresh. Primus's scheme is built around packing multiple encrypted inputs into a single bootstrapping and extracting multiple outputs from it, so one bootstrapping can absorb an entire cluster of gates rather than just one. Every cryptographic primitive in the stack is grounded in peer-reviewed, open literature, extended with original research from the Primus team.

Turning that idea into working circuits requires designing a custom encrypted gate for every logic pattern that appears, a task that used to require manual cryptographic expertise for each case. Primus's AutoHoG compiler automates this entirely: given any boolean function, it derives the gate's input weights and encrypted lookup table automatically, while proving the result stays within the noise budget. AutoHoG then scales this from a single gate to full circuits. It takes standard Verilog, runs it through industry-grade logic synthesis, and searches the resulting netlist for subcircuits that can be legally collapsed into a single custom gate, generating, validating, and substituting each gate automatically. Across adders, multipliers, and standard benchmark circuits, this compilation step alone cuts latency by up to 5.7x compared to naive gate-by-gate evaluation, before any hardware acceleration is applied.

The output of this compilation is remarkably uniform: regardless of the original circuit, the computation reduces to the same operation repeated many times — a weighted sum followed by a bootstrapping. Same memory pattern, same integer arithmetic, no data-dependent branching, which makes the workload a natural fit for GPU acceleration. Primus's GPU engine runs every polynomial operation in exact integer arithmetic via the number-theoretic transform (NTT), deliberately avoiding floating-point FFT, since floating-point rounding can drift across different hardware — unacceptable in a decentralized network where every node must arrive at the identical ciphertext. With pure integer math, independent nodes, even scheduled differently, reproduce byte-identical results across the full operator suite, giving the system consensus-grade determinism by construction.

Measured on 8x NVIDIA RTX 5090:

  • One bootstrapping: ~5 ms
  • One confidential transfer (64-bit encrypted balances: balance check, subtract, add, select): ~90 ms
  • Throughput: 210+ confidential transfers per second
  • Multi-GPU scaling: 8 GPUs deliver 8.0x the throughput of 1 GPU

Because the pipeline is latency-bound and clock-hungry, it runs efficiently on consumer hardware rather than requiring datacenter-exclusive accelerators, meaningfully lowering the cost of running confidential compute at scale. The same regular, integer-only, static-dataflow structure that makes this fast on GPUs also makes it a natural target for future FPGA and ASIC acceleration, a direction the scheme was designed for from the outset.