Niobium Developer

Understanding the Niobium FHE Software Development Kit

Overview

The Niobium FHE software development kit (SDK) is presently a collection of tools based on the openFHE library. The SDK is packaged in a Docker image. See Understanding the Niobium Client/Server RTS for more details.

Architecture

The overall architecture of the the Niobium FHE SDK is a library-oriented programming model. The application source code uses C/C++ function APIs to express the desired computation, and the library implements those operations.

This model is natural for normal CPUs, but is not well-suited for co-processor style architectures like GPUs and the Niobium FHE hardware accelerator.

For this reason, the current SDK should be considered experimental and provisional, and is under active evolution towards a more suited CUDA-like model.

Components

The core components to transform a C++ program into an instruction sequence for our hardware ISA (instruction set architecture) are:

  • The modified openFHE library containing C-probes at operation sites;
  • The C-probes library that implements collecting the operation sequences;
  • The custom compiler library that converts the operation sequences into ISA sequences and register allocations;
  • The FHE Softcore that processes the instruction sequence as a functional simulator and performance estimator.

If the source code being compiled has been separated into FHE client code and FHE server code, the FHE server code can be processed by the Niobium tool chain without modification.

If the client and server code are combined in a single source file, or group of source files compiled as a unit, the source code must be annotated to denote the region that the Niobium compiler should process to produce the instruction sequence.

The annotation has two parts. First, the Niobium compiler library header file needs to be included. The following example uses a conditional include based on a compile-time or source code definition to allow the code to used unmodified with either the Niobium tool chain or a normal C/C++ compiler:

#ifdef NIOBIUM_COMPILER
#include "niobium/compiler.hpp"
#endif

Second, the region where the FHE server code runs needs to be demarcated:

#ifdef NIOBIUM_COMPILER
  niobium::compiler().run(multiply, cc, ct1, ct2, multiplication);
#endif

In this example, the run function takes a C++ function-like reference, which may be a lambda.

Alternatively, the start and stop void functions can be used to demarcate the FHE server computation, and these functions can span the boundaries of other C/C++ functions.

Interactions

The interactions between the components in the SDK, for example, the compiler library, the modified openFHE library, etc. are all managed inside the Niobium runtime system Docker container.

These details are changing frequently and are not exposed.

Design Rationale

This design was the quickest and least-costly way to produce an instruction sequence for the Niobium FHE hardware accelerator ISA.

Constraints

The primary constraint driving the current design is the substantial overhead to rewrite the openFHE library in a way that is suitable for a compute co-processor architecture, like GPUs.

These limitations will start to be addressed by the LLVM-based experimental compiler, and a production compiler based on libraries or other representations of FHE computation (e.g. HEIR).

Extensibility

There is limited extensibility in the current design, other than the typical patterns applicable to the overall library-oriented application development paradigm.

Deployment

Deployment of the Niobium FHE SDK is fully contained in the production and distribution of the Niobium server RTS Docker image hosted on GitHub Container Registry.