ParaGato Labs · Edge genomics on RISC-V
Ginger-V is a research project exploring architecture- and memory-centric design space exploration for executing DNN-based genomic basecalling workloads on heterogeneous RISC-V systems targeting the edge.
The primary application focus is nanopore basecalling, where neural networks decode ionic current signals (“squiggles”) into DNA/RNA sequences under tight power, memory, and real-time throughput constraints.
Modern nanopore basecallers rely on deep neural networks and are typically
accelerated using GPUs to sustain real-time sequencing rates.
However, edge and field deployments demand alternatives that are:
- Energy-efficient
- Memory-aware
- Capable of sustained throughput matching signal generation rates
Key question:
When and how can genomic basecalling workloads be efficiently executed on FPGA-backed, heterogeneous RISC-V platforms?
Ginger-V studies this question through workload-driven architectural analysis, rather than proposing a new accelerator.
The project decomposes basecalling inference into:
- GEMM-dominant kernels
→ Offloaded to RedMulE, a systolic GEMM accelerator - Non-GEMM operations (control, activations, decoding, glue logic)
→ Executed on a RISC-V host core, with selective use of custom instructions
The emphasis is on:
- Memory traffic and data movement
- Tiling and reuse behavior
- Arithmetic intensity and throughput limits
- Real-time feasibility relative to nanopore signal rates
This repository leverages RedMulE, an outer-product systolic GEMM accelerator developed by the PULP platform, as the primary acceleration engine.
RedMulE is not original work of this project and is used strictly as an experimental substrate to study accelerator behavior and system integration.
Ginger-V is:
- A design space exploration framework
- A workload-driven architectural study
- A bridge between genomics workloads and reconfigurable hardware
Ginger-V is not:
- A new accelerator proposal
- A full ML compiler or runtime
- A production basecaller
RedMulE is included as a git submodule and points to a fork containing minor experimental and instrumentation changes used for analysis.
Original RedMulE repository:
https://github.com/pulp-platform/redmule