A large-array FPGA, a DDR3 interface, gigabit Ethernet and roughly 250 user I/O on one board — where the schematic is the easy part, and the power tree, the bank planning and the BGA breakout decide whether it works.
A data-acquisition board has a simple job to state and a hard one to build: bring a large number of external signals into an FPGA, hold them in fast memory, and move the data to a host. General-purpose, not welded to one converter or one experiment — which means the I/O count is high and the board has to stay honest across all of it.
The FPGA itself is the easy decision. A large Artix-7 — the XC7A200T in a 1156-ball package — has the logic and the pins. The engineering is everything around it: breaking out that BGA on a manufacturable stack-up, getting roughly two hundred and fifty single-ended I/O in and out without the board fighting itself, feeding a DDR3 interface and a gigabit Ethernet PHY, and powering the whole thing from a single external supply.
None of that is visible in the schematic. A netlist that is entirely correct can still produce a board that will not close timing, will not pass EMC, or will not power up in the right order — which is why this is a physical-layer programme, not a capture exercise.
Stated as the specification stated them, before any of them had an answer. A challenge described only after it was solved is a description of the solution.
Escaping every used ball of an FCBGA-1156 to routable layers sets the layer count and the stack-up. Get it wrong and the board is either unmanufacturable or unnecessarily expensive; both decisions are made before a single track is drawn.
Bringing about two hundred and fifty I/O out to headers — buffered through octal bus transceivers, with the bank and VCCO planning that implies — is where return paths and reference integrity are won or lost, not in the logic.
The FPGA and its interfaces need several rails at different voltages, sequenced in the right order and decoupled against a real PDN target. From one 5 V input, that is a tree of switchers, LDOs and a reference — a subsystem in its own right.
Memory fly-by topology and the Ethernet PHY have timing and layout rules that never appear on the schematic. They are constraints on the layout, and they have to be met on the same board as everything else.
The FPGA is the middle of the board, not the difficulty. The layer count is set by the BGA breakout; the schedule is set by the I/O fan-out, the DDR3 and Ethernet timing, and a power tree that has to come up in order from a single 5 V input.
The specific scope, rather than a capability list. This board was designed, laid out, built and brought up in-house.
Rarely the block that sounds difficult. Written out because a reader facing the same board gets more from this than from a summary of what went well.
Escaping a 1156-ball package decides the layer count before anything else is routed. That single decision propagates into cost, manufacturability and every timing budget on the board, so it was settled first.
The logic side of a large I/O count is trivial; the board side is not. Reference integrity, plane assignment and the buffering through octal transceivers are what keep the signals clean out to the headers.
Several rails from one input have to appear in the right sequence, hold their PDN target under the FPGA’s transient demand, and never latch the device into a bad state. The tree is a subsystem, and it was designed as one.
DDR3 fly-by and the gigabit Ethernet PHY carry layout rules that a correct schematic says nothing about. They are met on the board, on the same layers as everything else, or they are not met at all.
A large-array FPGA with roughly 250 user I/O available at headers, not tied to a single converter or experiment.
DDR3 for buffering, gigabit Ethernet and USB-UART for the host, brought up and validated rather than assumed.
Several rails sequenced from one 5 V input, decoupled to a PDN target and thermally managed for the FPGA.
Designed, laid out on a multilayer ENIG stack, assembled and functionally tested in-house — not handed off at netlist.
This programme is presented at property, capability and outcome level. The schematic, the full bill of materials, layout files and any customer identity are not published. Named silicon refers to commercially available third-party components. More detail is available under a non-disclosure agreement, within the limits agreed for the work.
Every item links to its own page, with the scope and maturity stated honestly for that item.
Schematic, layout, SI/PI, bring-up.
CAPABILITYArchitecture, HDL, closure, bring-up.
TECHNOLOGYDDR, Ethernet and the host paths.
CAPABILITYPower tree, references and integrity.
PRODUCTThe productised board family.
CAPABILITYBuild, assembly and test.
Not the FPGA. The difficulty is physical: escaping a large BGA on a manufacturable stack-up, keeping a high single-ended I/O count clean out to headers, meeting DDR3 and gigabit Ethernet timing in the layout, and building a sequenced multi-rail power tree from one input. A correct schematic guarantees none of it.
Buffering roughly 250 single-ended lines through octal transceivers protects the FPGA banks, sets a defined drive to the headers, and makes the VCCO and reference planning tractable — which is where signal integrity on a high-I/O board is actually decided.
The FPGA and its interfaces need several rails at different voltages, brought up in a defined order and each decoupled against a real power-distribution-network target. From a single 5 V input that is a designed subsystem of switchers, LDOs and a reference, not a handful of parts dropped in.
No. The board is presented at property and capability level; the schematic, bill of materials and layout are not published. Under a non-disclosure agreement, more can be discussed within the limits agreed for the work.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.