4 layers21 × 20 mm
4 layers167.7 × 111.2 mm
Lightrail ai gen3
Lightiam/-lightrail-ai-coreA Tiny Tapeout ASIC tile that forms the digital post-processor of the LightRail AI Gen3 LR-P8A photonic accelerator PCIe card.
- Size
- 167.7 × 111.2 mm
- Copper layers
- 4
- Thickness
- 2 mm
- Parts
- 11
- Nets
- 31
- Pads
- 2
- Tracks, vias
- 0, 0
- Routed
- Not yet
- KiCad
- KiCad 7 (file 20221018)
- Created
- 27 Mar 2026
- Updated
- 10 May 2026
Overview
Lightrail ai gen3 is an open-source printed circuit board by Lightiam, from the Lightiam/-lightrail-ai-core repository on GitHub, published under the Apache-2.0 licence. It is a 4-layer board of 167.7 × 111.2 mm with 11 parts on 31 nets.
Its parts include 3 ICs and 1 connector.
README
README.md
LightRail AI Gen3 – LR-P8A Photonic Inference Core
A Tiny Tapeout ASIC tile that forms the digital post-processor of the LightRail AI Gen3 LR-P8A photonic accelerator PCIe card.
What it does
Photonic AI accelerators perform matrix-vector multiplication at the speed of light using Mach-Zehnder Interferometer (MZI) meshes. After the optical computation, photodetectors convert results back to electrical signals and an ADC digitises them. This chip receives those 4-bit ADC readings and:
- Multiplies each reading by a signed 4-bit MZI weight (−8 to +7).
- Accumulates the product into one of four independent 8-bit signed channels (A, B, C, D) — mirroring the four fiber-optic input bundles on the LR-P8A card.
- Saturates at ±127/−128 instead of wrapping, matching the bounded optical power of a real photonic system.
- Applies ReLU (
max(0, acc)) on demand for neural-network activation.
Pin summary
| Pins | Signal | Description |
|---|---|---|
ui[3:0] | adc_in | Photodetector ADC result (unsigned 0–15) |
ui[7:4] | weight | MZI weight (signed −8…+7, 2's-complement) |
uio[1:0] | ch_sel | Channel select: 0=A 1=B 2=C 3=D |
uio[2] | relu_en | 1 = apply ReLU to output |
uio[3] | ch_clear | 1 = zero selected channel this cycle |
uo[7:0] | ch_out | Selected accumulator (post-ReLU if enabled) |
uio[7:4] | neg_d…neg_a | Sign flags – high when channel is negative |
What is Tiny Tapeout?
Tiny Tapeout is an educational project that makes it easier and cheaper than ever to get your digital designs manufactured on a real chip. Visit https://tinytapeout.com to learn more.
Resources
Parts
11| Qty | Value | Footprint | References |
|---|---|---|---|
| 2 | M3 | MountingHole:MountingHole_3.2mm_M3 | H1, H2 |
| 1 | LR-P8A-DIE1 | LR_P8A:ASIC_DIE_PH_40x40 | U1 |
| 1 | LR-P8A-DIE2 | LR_P8A:ASIC_DIE_PH_40x40 | U2 |
| 1 | PCIe_Retimer_Gen6 | LR_P8A:RETIMER_PH | U3 |
| 1 | PCIe_x16_Gen6 | LR_P8A:PCIE_X16_EDGE_PH | J1 |
| 1 | QSFP-DD_CH-A | LR_P8A:QSFP_DD_CAGE_PH | CAGE1 |
| 1 | QSFP-DD_CH-B | LR_P8A:QSFP_DD_CAGE_PH | CAGE2 |
| 1 | QSFP-DD_CH-C | LR_P8A:QSFP_DD_CAGE_PH | CAGE3 |
| 1 | QSFP-DD_CH-D | LR_P8A:QSFP_DD_CAGE_PH | CAGE4 |
| 1 | VRM_12V_TO_MULTI | LR_P8A:VRM_PH | VR1 |
Questions
How big is the Lightrail ai gen3 PCB?
It measures 167.7 × 111.2 mm and has 4 copper layers, 2 mm thick.
What parts does the Lightrail ai gen3 use?
11 parts. The full list of its parts is on this page, with links to each part's symbol and footprint where PCBKit's library has them.
Where can I download the Lightrail ai gen3 KiCad files?
From the Lightiam/-lightrail-ai-core repository on GitHub, at the commit this page was made from: the board is kicad/lightrail-ai-gen3.kicad_pcb and its schematic hardware/LR-P8A.kicad_sch. PCBKit opens the board in the browser, with nothing to install.
Can I use the Lightrail ai gen3 design in my own project?
Its repository publishes it under the Apache-2.0 licence: you may reuse it on that licence's terms (keep the author's notice, and for a copyleft licence share your changes alike).
Similar boards
- 4 layers47.9 × 47.9 mm
Flash2M FastBank
clydeshaffer/gametank
design files for the gametank hardware implementation
11 parts150 - 4 layers40 × 9 mm
UDB S
Unified-Daughterboard/UDB-S
Narrow variant unified daughterboards for custom keyboard designers and users
11 parts82MIT - 4 layers35 × 73 mm
Caps nightly
matthuszagh/fmcw
6GHz frequency-modulated continuous-wave radar with real-time range detection
11 parts66Apache-2.0 - 4 layers39.4 × 21.2 mm
- 4 layers22 × 22 mm
Caravel M.2-card WCSP
efabless/caravel_board
This repo provides firmware examples, flash programming and diagnostic tools for testing Open MPW and chipIgnite projects using Caravel.
11 parts40 - 4 layers71.3 × 59.7 mm
MemoryExpansion
RetroRemake/OtterX
8-Bit computer compatible with Commander X16
11 parts40CERN-OHL-W-2.0 - 4 layers33.3 × 88.1 mm
