T27.AI

Course

Train a neural network on an FPGA.

Not inference — training, on the chip itself. Eight modules from an empty toolchain to a network that learns on live silicon, entirely on open tools: no Vivado, no licences, and no step you cannot reproduce yourself.

Eight modules

01 · The open flow from nothing

Yosys, nextpnr-xilinx, prjxray, openFPGALoader and iverilog installed and verified on macOS arm64 or Linux. First bitstream blinking an LED on a real board, no vendor licence anywhere in the chain.

02 · Exactly as much Verilog as you need

Synchronous design, registers versus latches, and why an accidental latch is the classic bug that only shows up on silicon.

03 · Arithmetic as the foundation of ML in hardware

Why float is expensive, what quantisation really costs, and where ternary and low-precision formats come from.

04 · Bit-exact verification — the heart of the course

An independent Python reference model, per-stage known-answer vectors, checked through iverilog. Why a testbench written from the design's own assumptions cheerfully agrees with its bugs.

05 · A matrix multiplier that closes timing

MAC to array to pipeline. Reading the router's timing report and fighting for frequency on a real example — including why a hard multiplier in the path can leave you with no frequency report at all.

06 · Neural network inference on the FPGA

Layers, activations, dataflow and on-chip memory, running on the board rather than in a simulator.

07 · On-chip training — the capstone

Backward pass and SGD in RTL. The network learns XOR on the FPGA itself, 4 of 4, bit-exact against the reference. Almost nobody has done this by hand.

08 · Onward to silicon

The Tiny Tapeout path: preparing a design, what changes between FPGA and ASIC, and where the open silicon ecosystem stands after the move to IHP.

How this differs from the free alternatives

hls4ml (CERN) — free

Inference only, generated through HLS, and the flow underneath is a vendor toolchain. Excellent at what it does — it does not train on the chip, and it does not leave you able to read the RTL it produced.

Vendor courses (Intel, AMD) — free

Built to teach you their tools on their silicon. Nothing transfers to a flow you can run without a licence.

University FPGA courses

Usually stop at simulation, and where they reach a board it is through Vivado or Quartus.

What is left

Two things exist nowhere on that list: a backward pass running on the chip itself, and a flow with no vendor licence in it. If inference through HLS is what you need, use hls4ml — it is good, it is free, and I would tell you the same in an email.

Formats

Self-paced — $149

Video, code, KAT vector sets, community access.

Self-paced + hardware — $249

The same, plus remote runs on my Artix-7 boards. No board of your own required.

Cohort, 4 weeks — $599

Live sessions, code review, and your own design gone through with you.

Team workshop — from $2 000

Two days with your engineers around a problem you actually have.

Basic Python and some idea of digital logic is enough. Verilog is taught from scratch, and two of the formats include runs on my hardware.