dau. Investor overview

Draft

Items marked TBD have not been set. They live in src/lib/invest/overview.ts; the results come from src/lib/results.ts, shared with the home page.

Faster analytics, with an FPGA next to the CPU.

dau builds FPGA accelerators sized to analytical workloads, along with open-source software that decides how much of each query the FPGA should run.

01 · Thesis

Make the existing machine faster.

Analytical software already runs on fast CPUs. dau adds an FPGA card to the machine and splits the work between the two, so queries run faster without rewriting the software.

dau loads rows into device memory once and runs repeated filter, derive, and aggregate queries against them. A cost model picks the split for each query from measured rates. Anything the loaded hardware can't do runs on the CPU, and the planner reports it.

Device

Scans over resident rows.

Host

Sorts, text search, floating point, and everything else.

02 · What exists

Two boards running, five packages released.

Milestones on silicon
MilestoneStatus
TPC-H q1, q3, q6, q19Bit-exact on dpv1 silicon
Full-day TAQ to OHLCV bars137.4M rows, every bar equal to Polars
Rebuild-and-flash loopWorkload to new bitstream to silicon, bit-exact
Four execution posturesCPU, offload, split, adaptive; 8 of 8 bit-exact
macOS hostFull DMA from a Mac over Thunderbolt
dpv2 bring-upTPC-H q6 bit-exact on both boards
Software
PackageReleaseScope
dau-core0.4.0 on PyPIOperator tiles, golden models, capability contracts
dau-build0.15.2 on PyPISpecs to bitstream, Vivado or yosys
dau-sim0.3.0 on PyPICycle-accurate simulation
dau-driver0.1.0 on PyPIXDMA host I/O on Linux and macOS
dau-scheduler0.1.0 on PyPICost-based host/device split
dau-polarsPrivate, release in preparationPolars plan capture and execution

dpv1

First platform

Thunderbolt device

Specifications
ParameterValue
Logic134K LUTs
Memory1 GB DDR3
Host linkPCIe Gen2 x2 over Thunderbolt

dpv2

Current platform

PCIe card

Specifications
ParameterValue
Logic204K LUTs
Memory2 GB onboard + 8 GB SODIMM DDR3
Host linkPCIe x8
Card costAbout $700

dpv3

In design

PCIe card for servers

Specifications
ParameterValue
MemoryDDR4, capacity to be set
Host linkPCIe
Card to cardDirect fabric links

03 · Evidence

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. Every figure matches its software reference exactly.

Benchmark results on dau boards
BenchmarkBoardResult
TPC-H query 6, with the table loaded on the carddpv2

~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 barsdpv1

137.4M trades, 2.44M bars

Every bar matches the CPU Polars result.

Time-series workflow: as-of join, derived features, time bars, rolling statisticsdpv2

1,024,064 records in 11.5 ms

All four steps run in one pass, and intermediate results stay on the card.

04 · Business model

Open core, dau boards.

The software is Apache 2.0, published on PyPI, and runs on open FPGA hardware. Revenue comes from boards built for dau and from configurations we build and verify for a customer's workloads.

ParameterValue
PricingTBD

05 · Market

Repeated queries over the same data.

The first workloads are time-series and market data: bars, windows, and as-of joins run again and again over the same history.

ParameterValue
Market sizingTBD
First customersTBD

06 · Roadmap

Faster links first, then more cards.

  1. Release the Polars integration so the whole stack installs from PyPI.
  2. Design dpv3 around the measured bottleneck, memory bandwidth and the host link, with an FPGA chosen for this workload.
  3. Spread work across cards with direct links between cards and memory on each.

07 · Risk

Risks.

Host CPUs keep getting faster
dau has to keep adding speed on top of each new generation of host CPU, which affects which workloads it targets.
Data transfer is the bottleneck
On dpv1 and dpv2, moving data to the card takes longer than processing it. Keeping data on the card and faster links on dpv3 address this. Until then, the best fit is workloads that query the same data repeatedly.
Toolchain dependence
The build depends on AMD Vivado. Using an FPGA outside that toolchain means porting the build, however cheap the part.
dpv3 part selection is open
The production card is not designed yet, so its cost, memory, and lead times are not known.

08 · The ask

The raise.

ParameterValue
AmountTBD
InstrumentTBD
ValuationTBD
Use of funds
ItemAmount
dpv3 design and first buildTBD
Engineering hiresTBD
Design-partner deploymentsTBD

Questions and data-room requests: admin@1kbgz.com