Personal
Next to your laptop
A Thunderbolt device connected to your laptop.
Data, accelerated.
dau runs selected Polars, dataframe, and time-series operations as FPGA dataflow and leaves the rest in software. The same software works with an accelerator next to your laptop or a card in a server.
How a query is split
Hardware
Scans run on the FPGA.
Filters, derived columns, and aggregations stream through tiles built for the workload. Rows can stay in device memory between queries.
Software
Everything else runs in Polars.
If the loaded configuration doesn't cover an operation, it runs on the CPU, and the planner reports what ran where.
Platform
Hardware built for your workload, with the software you already use.
Configurations are assembled from reusable operations and sized to the job. Your code stays the same whether it runs on one accelerator next to a laptop or several cards in a server.
Scale-out
Several on one laptop
Several Thunderbolt devices connected to one laptop, with work split across them.
Workstation
In your desktop
A PCIe card in your desktop.
Planned
In a server
Several PCIe cards in a server, linked directly to each other.
Each deployment shapes the next design.
We measure real workloads to decide what runs in hardware now and what to build next. Every design is verified the same way before it ships.
- 01 Observe Profile real workloads.
- 02 Select Choose the operations worth accelerating.
- 03 Compose Build a right-sized configuration from tested tiles.
- 04 Verify Compare against the software reference in simulation and on the board.
- 05 ↺ Measure Deploy, measure, and use the results in the next design.
Results
About 2.5× faster with a dau card.
*On TPC-H query 6, a standard analytics benchmark, adding a dau card to the host made the query about 2.5× faster. The table below lists benchmarks we have run on our boards. Every result matches the software reference exactly.
| Benchmark | Board | Result |
|---|---|---|
| TPC-H query 6, with the table loaded on the card | dpv2 | ~2.5× faster with the card added With a six-million-row table on the card, the CPU and card together finished in 4.2 ms, against 10.4 ms for the CPU alone. The planner picks the split from measured costs. |
| A full day of NYSE TAQ trades turned into OHLCV bars | dpv1 | 137.4M trades, 2.44M bars Every bar matches the CPU Polars result. |
| Time-series workflow: as-of join, derived features, time bars, rolling statistics | dpv2 | 1,024,064 records in 11.5 ms All four steps run in one pass, and intermediate results stay on the card. |
dpv1
First platformThunderbolt device
Specifications
| Parameter | Value |
|---|---|
| Logic | 134K LUTs |
| Memory | 1 GB DDR3 |
| Host link | PCIe Gen2 x2 over Thunderbolt |
dpv2
Current platformPCIe card
Specifications
| Parameter | Value |
|---|---|
| Logic | 204K LUTs |
| Memory | 2 GB onboard + 8 GB SODIMM DDR3 |
| Host link | PCIe x8 |
| Card cost | About $700 |
dpv3
In designPCIe card for servers
Specifications
| Parameter | Value |
|---|---|
| Memory | DDR4, capacity to be set |
| Host link | PCIe |
| Card to card | Direct fabric links |
Open source
An open core that runs on open hardware.
The operator tiles and their reference models, the build system, the simulator, the host driver, and the scheduler are Apache 2.0 licensed and published on GitHub and PyPI. They run on open FPGA hardware, so you can build the same bitstreams we run and add tiles in your own package. Boards built for dau are available through the design partner program. The Polars integration will be released soon.
pip install dau-core dau-driver dau-scheduler- dau-core Operator tiles in SystemVerilog, with bit-exact reference models, capability contracts, and the registry.
- dau-driver Host I/O over XDMA on Linux and macOS: discovery, registers, DMA, and jobs.
- dau-scheduler Decides how much of each query runs on the FPGA, based on measured costs.
- dau-build Declarative build specs through synthesis, programming, and smoke tests.
- dau-sim Cycle-accurate simulation for Amaranth and SystemVerilog, with a cocotb backend.
- All repositories on GitHub →
Design partners
Work with us.
We're working with early users on dau hardware, software, and deployment. Tell us what you compute, where it runs, and how fast it needs to be.