Access Official Technical PDF

Share Summary:LinkedInEmail

FPGA probabilistic AI offers one way to accelerate workloads that rely on repeated probability calculations rather than conventional deterministic processing. A research design presented at the 2019 IEEE International Conference on Rebooting Computing applied this approach to gene network modeling and personalized breast cancer treatment. Its broader engineering lesson is how to divide a complex AI workload between software control and reconfigurable hardware—while keeping the model interpretable and its results appropriately qualified.

Why use probabilistic models for gene networks?

Genes and proteins interact in networks, and changes in those interactions can be associated with disease. A Bayesian network (BN) represents entities as variables and their relationships as directed connections. Conditional probability tables (CPTs) describe how the probability of one variable’s state depends on the states of its parent variables.

This structure supports two computational tasks. Learning estimates the network structure and its probability parameters from data, sometimes alongside prior knowledge from domain experts. Inference uses the network to estimate probabilities given observed information. In the research application, gene expression data helped build gene networks and informed a model that estimated the probability of cancer recurrence after treatment.

Exact inference can be computationally demanding. The design therefore used Gibbs sampling, a sampling-based method that estimates probabilities through repeated samples. Sampling is approximate, but its operations can be carried out in parallel, making it a candidate for hardware acceleration.

FPGA probabilistic AI architecture: software and hardware roles

The system was implemented on a system-on-chip FPGA platform, which combines a processor subsystem with programmable FPGA logic. Rather than placing the entire workflow in custom hardware, the design assigned sequential control tasks to software and parallelizable calculations to the FPGA.

The processor handled candidate network generation and selection. It added genes in a predetermined order, formed candidate connections, and compared scores to decide which structures to retain. The FPGA evaluated candidate structures by performing the repeated sampling operations used in scoring. It was also used for inference on the learned network to calculate a recurrence score.

Configurable sampling units

On the FPGA, network nodes were mapped to Gibbs Sampling Units. Each unit stored CPT values and used a linear feedback shift register (LFSR) as a pseudo-random number generator. A node’s parent state selected a CPT entry; the resulting probability was compared with a random value to determine the node’s sampled state.

Units could be connected through a programmable crossbar interconnect, allowing the hardware to represent different network structures. Software configured parameters such as connectivity, CPT values and the number of samples through memory-mapped registers. This separation let the processor manage the search while the FPGA performed sampling-intensive work.

From gene expression to a recurrence score

The treatment model used a hierarchy of variables. Observed gene expression values informed intermediate, or latent, variables representing cancer-related subscores. Those subscores fed a higher-level estimate of recurrence probability. The model’s output was intended to support treatment selection by estimating whether recurrence risk was low or high.

The study evaluated the approach using expression data for 3,070 breast cancer patients, with information on cancer type, treatment and survival events. The authors reported that model-generated treatment suggestions agreed with recorded physician choices in about 80% of cases. The remaining differences indicate that the model and the treatment decisions in the dataset did not always align; they do not, on their own, establish that a model recommendation would be clinically preferable.

That distinction matters for technical decision-makers. Agreement with historical decisions is a comparison, not proof of clinical effectiveness or readiness for deployment. The work describes a computational research approach, not a substitute for clinician judgment or clinical validation.

What the hardware results suggest

The paper reports that its SoC-FPGA implementation ran 25 times faster than a software implementation on a 16-core Xeon workstation while consuming 25 times less power. These figures describe the reported implementation and comparison; they should not be treated as a general performance guarantee for other models, devices or workloads.

Resource use also increased with network size. On the Cyclone V FPGA, the reported design used 14% of available resources for an 18-node network and 41% for a 41-node network. The authors estimated that networks with hundreds of nodes could be feasible, subject to FPGA logic availability. For smaller graphs, multiple candidates could potentially be evaluated concurrently, subject to the available hardware resources.

For system architects, the design illustrates a useful pattern: identify the repetitive, parallel portion of an algorithm, map it to specialized hardware, and retain software for flexible control and model management. Reconfigurability is particularly relevant when network connections or model parameters need to change.

Practical implications for automation and edge computing

Although the application is medical, the architecture raises questions relevant to industrial computing. Probabilistic models may involve intensive sampling, while deployed systems also need configurable logic, a clear division of responsibilities between processor and accelerator, and careful management of model inputs and outputs.

When evaluating a similar design, engineers and integrators should consider:

  • Workload fit: Determine whether the dominant calculations can be parallelized and are suitable for FPGA implementation.
  • Model flexibility: Assess how often network structure, parameters or data formats may change, and how those changes will be configured.
  • Resource scaling: Measure logic use as the model grows rather than assuming a small demonstration will scale without constraint.
  • Validation: Separate computational performance from evidence that model outputs are accurate and appropriate for their intended use.

Conclusion

This FPGA probabilistic AI design shows how Bayesian network learning and inference can be divided across a processor and reconfigurable hardware. Its sampling units and programmable interconnect target parallel probability calculations, while software manages candidate generation and selection. The reported results point to the potential of specialized acceleration, but also underline the need to assess resource growth and validate application outcomes independently. For engineering teams, the key takeaway is not simply to move AI onto an FPGA, but to match the hardware architecture to the workload and define clearly what the model’s results can—and cannot—support.

Access Official Technical PDF

About The Author