Data. Accelerated.
by 1kbgz.

dau.dev · Python API · Polars frontend · Build tools · Simulation

DAU is a hardware/software stack for accelerating analytical queries. It maps SQL-style operators onto a reconfigurable, tile-based FPGA dataflow engine, moves fixed-width Arrow record batches over PCIe, and drives execution from Python dataframe workflows.

The long-term shape is a database processing unit: host-analyzed query plans, composable operator tiles, a host-configured on-chip network, explicit CPU fallback for unsupported plan fragments, and repeatable build/simulation flows for FPGA targets.

• Frontend: lazy Polars integration for selective pushdown of supported operators.
• Runtime: Python APIs, device discovery, register access, DMA helpers, Arrow-derived stream codecs, and result rematerialization.
• Hardware: reusable SystemVerilog operator tiles with golden software models, cocotb benches, and Verilator benches.
• Build flow: declarative specs, artifact manifests, Vivado/XDMA handoff, simulation, synthesis, flash, and smoke-test task surfaces.
• Current status: internal end-to-end market-data aggregation workloads have run on FPGA and matched CPU goldens.
• Current focus: broader operator coverage, runtime scheduling, result streaming, and throughput work.

• dau (private) - thin end-user Python API over stable DAU primitives.
• dau-polars (private) - Polars frontend for selective FPGA pushdown with CPU fallback.
• dau-core (private) - hardware-facing contracts, golden semantics, stream protocols, and reusable HDL.
• dau-driver (private) - DAU-compatible device discovery, registers, DMA, codecs, and execution helpers.
• dau-build (public) - build specs, artifact bundles, generated hardware handoff, and task orchestration.
• dau-sim (public) - simulation infrastructure for digital hardware designs, including cocotb and Verilator integration.
• dau-utils (private) - shared host utilities used by the stack.
• artlink (public) - domain-neutral artifact manifests, validation templates, and registry/discovery helpers.

Some DAU development is proprietary or hardware-lab specific, including portions of the accelerator implementation, bring-up evidence, datasets, and internal integration notes. Open-source DAU code is released under the Apache 2.0 License unless a repository states otherwise.