I2C, SPI and UART IP Cores: The Glue Logic Every FPGA Needs

I2C, SPI and UART IP Cores: The Glue Logic Every FPGA Needs

Industry news — On 14 July 2026 the MIPI Alliance’s I/O Bridges Working Group demonstrated two independent implementations of its forthcoming Specification for I/Os Over I3C Bridge, using Microchip, Binho and NXP parts to reach I2C, SPI, UART and GPIO devices through a single I3C controller — the industry is consolidating access to the legacy serial interfaces, not retiring them (MIPI Alliance, 14 July 2026).

In short: the I2C, SPI and UART IP cores in an FPGA design are small, unglamorous and routinely the reason integration slips — because they are the only blocks that talk to parts you did not design

Read that demonstration carefully and it says something useful about the next ten years of board design: the body standardising the future of on-board signalling spent its engineering effort building a bridge to I2C, SPI and UART, not a replacement for them. The endpoints are not going anywhere. The SFP cage still answers on a two-wire bus, the configuration flash still wants a chip select, and the maintenance connector on the front panel is still a UART. An FPGA sitting in the middle of that board has to speak all three natively, in fabric, and it is almost always these small macros — not the NVMe or Ethernet core everyone budgeted for — that decide when the board actually integrates.

The short answer

I2C, SPI and UART are the three low-rate serial interfaces almost every FPGA board needs, and they are cheap to underestimate. Logic Design Solutions publishes a family of small VHDL macros that cover them: MI2CM (I2C master), MI2CS (I2C slave) and MI2CMS (both) — compliant with the I2C-bus specification version 1.0, standard mode at 100 Kbits, 7-bit and 10-bit addressing, hardware digital filtering on SCL and SDA, no general call; MSPIM (SPI master, up to 8 slaves) and MSPIS (SPI slave, system clock up to 220 MHz, extended mode with 14-bit address and 16 or 32 bit data); and three UARTs — MUART, M16550 (based on the National Semiconductor NS16550, up to 1M baud) and M16450 (system clock up to 150 MHz). MUART, MSPIM and MSPIS are published with a DO254 designation. All are delivered as VHDL source and can be customised by Logic Design Solutions.

Why the unglamorous cores decide the schedule

A storage or avionics FPGA programme is rarely late because of the NVMe core. That block is the one everybody stared at during the architecture review: it had a schedule line, a verification plan, an evaluation board and a named owner. The blocks that slip are the ones nobody assigned, because everybody assumed they were free.

Glue logic is the old name for that category. On a modern FPGA board it is not really glue any more, it is the management plane: reading the identification EEPROM out of an SFP module before you dare bring the link up, polling a temperature sensor to decide whether to throttle, loading a calibration table from serial flash at boot, answering a technician on a maintenance port when something goes wrong in the field.

Three things make these interfaces expensive when improvised rather than licensed. They are the only blocks that touch parts you did not choose — an optical module, a sensor, a supplier’s sub-assembly — so they are where specification mismatches surface. They fail in ways that look like something else: a bus that wedges once a day reads as a mysterious system hang, not as an I2C problem. And they are exercised last, because they only become interesting once real hardware exists.

I2C, SPI and UART compared

The three interfaces are not alternatives to one another — a real board carries all three, each doing the job it is shaped for. The table sets out the differences that decide which one a peripheral appears on, and which LDS macro covers it.

I2C, SPI and UART on an FPGA board
Criterion I2C SPI UART
Signals Two, SCL and SDA, open-drain with pull-ups Four for one device: SCLK, MOSI, MISO, chip select — one more select per device Two, TX and RX; optional RTS/CTS flow control
Addressing In-band device address, 7-bit or 10-bit None — the chip select is the address None — the link is point-to-point
Clocking Synchronous; master drives SCL, a slow slave may stretch it Synchronous; master drives SCLK, polarity and phase programmable Asynchronous; no clock line, both ends agree a baud rate
Rate 100 kbit/s standard mode; the specification defines fast, fast-plus, high-speed and ultra-fast tiers up to 5 Mbit/s No specified ceiling — set by the FPGA, the board and the slave Typically 9.6 to 115.2 kbaud, up to 1M baud on M16550 and M16450
Multi-drop Yes — many devices share one pair; multi-master with arbitration No; a star of chip selects, or a vendor-specific daisy chain No; RS-485 adds it at the transceiver, not at the UART
Error detection Per-byte acknowledge or not-acknowledge; no CRC None in the interface; any check is device-specific Parity, framing, break and overrun, flagged per byte
Typical FPGA-board use SFP and QSFP module management, temperature sensors, EEPROMs, clock generators, power rails Configuration and calibration flash, ADCs and DACs, RF transceivers, displays Maintenance and console ports, GNSS receivers, legacy industrial and avionics kit
LDS macro MI2CM, MI2CS, MI2CMS MSPIM, MSPIS MUART, M16550, M16450

Two patterns fall out. Anything that hangs several cheap devices off two pins goes on I2C and accepts the low rate. Anything needing bandwidth to a single device, with pins to spare, goes on SPI. Anything that has to be readable by a human with a terminal stays on a UART.

I2C: what the macros do, and what they deliberately do not

MI2CM is master-only, MI2CS is slave-only, and MI2CMS is the combined macro. All three carry the same published feature set: compliance with the I2C-bus specification version 1.0, “Standard mode operation (100Kbits)”, reads and writes only, 7-bit and 10-bit address management, synchronisation on the system clock, a hardware digital filter on SCL and SDA, no general call support, and a system clock up to 180 MHz depending on FPGA speed grade. The master-capable variants add multi-master operation with arbitration and clock synchronisation.

What “standard mode only” costs you

The I2C-bus specification, maintained by NXP as UM10204, defines several bit-rate tiers: 100 kbit/s in standard mode, 400 kbit/s in fast mode, 1 Mbit/s in fast-mode plus, 3.4 Mbit/s in high-speed mode and a unidirectional ultra-fast mode reaching 5 Mbit/s. A standard-mode macro sits at the bottom of that ladder, and for the management plane that is almost always the right place to be: an SFP identification page is a few hundred bytes, a temperature register is two. Where it bites is any design that treats I2C as a data path rather than a control path — firmware pushed over the bus, camera modules configured frame by frame. If your budget assumes 400 kbit/s, that is a customisation conversation to have before the schematic is frozen, not after.

What “no general call” really means

The general call is the I2C broadcast: address 0000 000 with the read/write bit low addresses every device at once, and the second byte says what the broadcast means — 06h resets all responding devices and reloads the programmable part of their address, 04h does the same without the reset. It is also optional. UM10204’s own table of protocol features lists the general call address, 10-bit addressing, clock stretching and software reset as optional in every configuration, while start and stop conditions, acknowledge and 7-bit addressing are mandatory. The omission is therefore a documented scope decision, not a compliance gap. Practically: on the slave side a broadcast to address 0x00 will not be acknowledged, so a general-call software reset is not available as a bus-recovery path; on the master side the macro will not originate one, so devices that expect collective addressing must be addressed one at a time.

Why a hardware filter on SCL and SDA earns its logic

I2C is an open-drain bus: nothing drives high, a pull-up resistor does, and every edge is an RC curve rather than a switched transition. That makes it easy to route across mixed voltages, and permanently susceptible to anything coupling into it — a switching regulator a centimetre away, a long backplane run, a ribbon cable in a rack full of motors. The consequence of a glitch is asymmetric, and that is the part worth internalising. A spike on SDA corrupts one bit, which the transaction may survive. A spike on SCL is read as an extra clock edge, and from that moment master and slave disagree about which bit of the byte they are on. The classic end state is a slave that thinks it is mid-read, holding SDA low indefinitely, with the bus wedged for everyone on it — recoverable only by toggling clock pulses until it releases, or by a power cycle. On a sealed box in an aircraft bay that is not a debug session, it is a maintenance event. UM10204 requires input filters on SDA and SCL that suppress noise spikes shorter than 50 ns on fast-mode and fast-mode-plus devices; the LDS macros publish “Hardware digital filter on SCL and SDA signals” as a feature in its own right — inside the macro, in VHDL you can read, rather than an RC network improvised on two nets at board level.

SPI: four modes, no standards body

SPI is the odd one out, because there is no SPI standard. No standards body, no revision number, nothing to be compliant with: it began as a Motorola on-chip peripheral interface and became universal by being copied, so the authoritative document for any SPI link is the datasheet of the device at the far end. Compare that with I2C, where a numbered specification says what a compliant device must do, or with ARINC 429, where an aviation standard defines the word format down to the bit. On SPI the interoperability risk sits entirely with the integrator.

The parameters that have to match are CPOL and CPHA — clock polarity and clock phase. CPOL says whether the clock idles low or high; CPHA says whether data is sampled on the first edge of a bit cell or the second. The four combinations are the four SPI modes, numbered 0 to 3, and a mismatch does not produce a clean error: it produces data shifted by half a bit period, which reads back as plausible nonsense. Both MSPIM and MSPIS publish a serial clock “programmable with polarity and phase”, which is what lets one macro talk to devices that disagree about the convention.

Each then adds two run-time modes. Standard SPI mode implements bidirectional one-byte transactions. Extended SPI mode implements frame transactions: on MSPIM as an SPI EEPROM controller, on MSPIS as frames carrying a 14-bit address and 16 or 32 bits of data — worth reading twice if you are building a register interface between two FPGAs, because it turns SPI from a byte pipe into an addressed transaction without a framing layer of your own. The published rate constraints differ: MSPIS gives a system clock ceiling of 220 MHz, speed-grade dependent, while MSPIM publishes no clock figure but states that the serial clock period can be at least six system clock periods — a ratio rather than a number, and the honest way to size SCK before synthesis. MSPIM supports up to eight SPI slaves.

UART: why NS16550 compatibility still buys you a driver

A UART is the only one of the three interfaces that carries no clock line, and that single fact explains its whole design. Transmitter and receiver agree a baud rate in advance, a start bit announces each character, and the receiver rebuilds the timing locally. The LDS UARTs do it the classical way: a divisor of 1 to 65535 generates a 16X clock, so the receiver oversamples each bit sixteen times and decides in the middle of the bit cell, where a small baud-rate error between the two ends does least damage.

All three implement the same asynchronous protocol: 5, 6, 7 or 8 bits of data, even, odd or no parity, start and stop bit generation and detection, line break generation and detection, receiver overrun and framing detection. MUART and M16550 buffer transmit and receive with a FIFO 16 bytes deep, plus three error bits per data byte on the receiver, with polled or interrupt operation and a loopback mode; MUART adds a register mode alongside FIFO mode. M16450 is the lighter one — buffered transmit and receive registers, no FIFO in its published feature list, a system clock up to 150 MHz, up to 1M baud — which is the lineage split you would expect, since the FIFO is what the 16550 generation added to the 16450 generation.

The register map is the product

M16550 is described on its page as “Functionally based on the National Semiconductor Corporation NS16550 device”, and that sentence is the reason to choose it over a cleaner UART you could write yourself. The NS16550 register layout — receive and transmit holding registers, interrupt enable, interrupt identification and FIFO control, line control, modem control, line status and modem status, with the divisor latches behind a bit in the line control register — became the universal serial port interface. Linux carries an 8250/16550 driver maintained for decades, and boot loaders, real-time operating systems and firmware consoles assume that layout. Expose an M16550 at a sensible base address and existing software binds with a device-tree entry or a base-address constant. Present a bespoke UART and you have added a driver to the software backlog and code to maintain for the life of the product — the same trade that runs through what an FPGA IP core actually is.

What the DO254 designation actually means

DO-254 is the RTCA document titled Design Assurance Guidance for Airborne Electronic Hardware, and it defines a process — requirements capture, design data, verification evidence, configuration management, traceability from requirement to test — for complex electronic hardware in certified aircraft. Civil aviation authorities on both sides of the Atlantic recognise it as an acceptable means of compliance, and it applies to design assurance levels from A to E, assigned by the safety assessment of the aircraft function the hardware performs.

Four products in this family carry the designation in their name: MUART, MSPIM, MSPIS and the legacy ARINC 429 macro M429T1R1. The wording is specific — MSPIS and M429T1R1 offer “DO254 design, verification and traceability documentation available on request”, while MUART and MSPIM state that DO254 documentation is available. The same phrasing appears elsewhere in the catalogue, for example on the current ARINC 429 IP. The three I2C macros do not carry it on their published pages.

What that tells an avionics integrator is that the design data exists in a form an applicant can fold into a certification package: requirements traced to design, verification traced to requirements, configuration control over the set. Reconstructing those artefacts after the fact is expensive and sometimes impossible, so their existence is a real procurement criterion. What it does not tell you is a design assurance level. LDS publishes no DAL for these macros and none should be inferred — and in any case a DAL attaches to an item in an aircraft, not to an IP block in a catalogue: the applicant certifies the equipment, the supplier supplies evidence into that effort. If you need a specific level, ask with your certification plan in hand, alongside the rest of the due diligence any IP vendor should survive.

VHDL source, and why it matters more here

Every macro in this family is offered “in VHDL source code format for ease of customization”, with the added line that it “can be customised by Logic Design Solutions”. Source delivery is a differentiator on any IP, but on a peripheral macro it is worth more than on a large protocol core, for a simple reason: large cores are used as specified, and small ones are modified.

Nobody asks for a variant NVMe host that queues commands differently — the specification is the point. Peripheral macros are the opposite, and the requests are mundane and endless: bring the CPU-side interface out as AXI4-Lite instead of the native register set, widen a FIFO because a burst of console traffic overruns it, add a bus-recovery sequence that clocks a wedged I2C slave free, decode a chip select for nine devices instead of eight, timestamp received UART characters. Each is a small change against a known design, impossible against an encrypted netlist. Source also underwrites the certification argument above, since a DO-254 evidence package is only meaningful if the design data it traces to is the design data you compile — which is why this family turns up in design-services engagements as often as in catalogue licensing.

Integrating the macros into a busy FPGA

Each macro exposes a direct connection to a CPU register set, and that is the first integration decision. On a device that already carries a processor — a Zynq UltraScale+ from AMD, an Agilex or Arria part from Altera, a PolarFire SoC from Microchip — the natural move is to wrap the native register interface in the fabric bus the design already uses, so the peripheral lands in the same address map as everything else. With no processor, the same registers are driven by a state machine in fabric — the macros work either way, because they were written as CPU-independent logic.

Three details are worth planning rather than discovering. Interrupts: all three UARTs offer polled or interrupt operation, and on a console that also carries telemetry that choice decides whether the FIFO matters. Clock domains: every macro is synchronised on the system clock, with published ceilings of 180 MHz for the I2C macros, 220 MHz for MSPIS and 150 MHz for M16450 — comfortable numbers, but speed-grade dependent, so a constraint to confirm rather than assume. Pin planning: I2C needs two open-drain-capable pins with correctly sized pull-ups, SPI four plus one select per device, a UART two — and pin assignment gets locked at schematic freeze, months before anyone writes a line of VHDL.

These macros belong to the same board as the fast cores, not to a different project. The 10G Ethernet MAC carrying your data plane sits behind an SFP cage whose module you can only identify over I2C, and the NVMe host core writing your recordings shares a device with the maintenance port that will explain, one day, why it stopped.

Tell us the peripherals on your board, the target FPGA family and whether the programme is heading for certification, and we will tell you which macro fits as published and what would have to change.

Talk to an FPGA engineer

Where this stops applying

This article describes what Logic Design Solutions publishes about these macros. Several things it deliberately does not do:

  • No resource utilisation, power or latency figures. LDS does not publish LUT, flip-flop, block-RAM or DSP counts, power consumption or transaction latency for any macro in this family. Anyone quoting you such a number for these cores is estimating. Ask the supplier for figures against your target device and speed grade.
  • No design assurance level. The DO254 designation on MUART, MSPIM, MSPIS and M429T1R1 means documentation is offered; it is not a published DAL, and this article does not imply one.
  • Published clock ceilings are speed-grade dependent. The 180 MHz, 220 MHz and 150 MHz figures come from the product pages and are qualified there as depending on FPGA speed grade. They are a starting point for sizing, not a timing sign-off.
  • Standard mode is a real constraint on I2C. If your bus budget assumes fast mode or faster, or your architecture depends on general call, these macros as published do not cover it — that becomes a customisation conversation, not a data-sheet lookup.
  • SPI interoperability cannot be verified from a specification. Because SPI has no standards body, the only proof that a master and a slave agree is the far-end datasheet and a scope trace on real hardware.

In summary

The serial peripheral macros are the least glamorous line in an FPGA IP catalogue and the one most likely to determine when a board integrates. The LDS family covers all three interfaces: I2C through MI2CM, MI2CS and MI2CMS, with standard-mode operation, 7-bit and 10-bit addressing, multi-master arbitration and a hardware filter on SCL and SDA; SPI through MSPIM and MSPIS, with programmable polarity and phase and an extended frame mode; and UART through MUART, M16550 and M16450, where NS16550 compatibility means an existing driver stack binds without new software.

What separates them from something written in-house in a fortnight is not cleverness. It is that the awkward cases — a glitch on SCL, a phase mismatch on SCK, an overrun on a console under load — have already been found and fixed, the design arrives as readable VHDL, and on the DO254-designated macros the evidence trail an avionics applicant needs already exists.

Frequently asked questions

What is the difference between MI2CM, MI2CS and MI2CMS?

They are the three I2C variants of the same macro family. MI2CM implements the master function only, MI2CS the slave function only, and MI2CMS both in one block. All three publish the same core feature set — compliance with the I2C-bus specification version 1.0, standard mode at 100 Kbits, 7-bit and 10-bit addressing, a hardware digital filter on SCL and SDA, no general call support, and a system clock up to 180 MHz depending on speed grade. The master-capable variants add multi-master operation with arbitration and clock synchronisation.

Why does no general call support matter, and when does it not?

General call is the I2C broadcast: address 0000 000 addresses every device on the bus at once, with the second byte defining the action — 06h resets responding devices and reloads the programmable part of their address, 04h does so without a reset. The I2C-bus specification lists it as optional in every configuration, so omitting it is a scope decision rather than a compliance gap. It matters only if your architecture uses broadcast reset or broadcast addressing; on most boards, devices are addressed individually and nothing is lost.

How fast can these SPI macros run?

The published figures differ by macro. MSPIS gives a system clock ceiling of 220 MHz, qualified as FPGA speed-grade dependent. MSPIM does not publish a system clock figure but states that the serial clock period can be at least six system clock periods, which caps SCK at a sixth of whatever system clock you close timing at. LDS publishes no measured throughput for either macro, so any figure in megabytes per second would have to be derived for your device and confirmed on hardware.

Is 16550 compatibility still worth anything in a new design?

Yes, and the value is in software rather than hardware. The NS16550 register layout is what serial drivers across Linux, real-time operating systems, boot loaders and firmware consoles expect. M16550 is described as functionally based on the NS16550 device, so an existing driver binds to it with an address and an interrupt rather than with new code. A bespoke UART, however well designed, adds a driver to write, debug and maintain for the life of the product.

Does the DO254 designation mean these cores are certified?

No. DO-254 is design assurance guidance for airborne electronic hardware, and a design assurance level attaches to equipment in an aircraft rather than to an IP block in a catalogue. MUART, MSPIM, MSPIS and M429T1R1 are published with the designation, and MSPIS and M429T1R1 specifically offer design, verification and traceability documentation on request. That means evidence exists for an applicant to fold into a certification package. LDS does not publish a DAL for these macros, and none should be inferred.

Can these macros be modified for a non-standard requirement?

That is the stated model. Every macro in the family is offered in VHDL source code format for ease of customisation, with the explicit note that it can be customised by Logic Design Solutions. On peripheral logic that matters more than on a large protocol core, because the requests are constant and small: a different CPU-side bus, a deeper FIFO, an extra chip select, a bus-recovery sequence, a different interrupt style. Those are routine changes against readable source and impossible against an encrypted netlist.

  1. NXP Semiconductors — I2C-bus specification and user manual (UM10204), Rev. 7.0, 1 October 2021: bit-rate tiers, reserved addresses, the general call second byte, the optionality table and the 50 ns input-filter requirement.
  2. MIPI Alliance — Working group demonstration highlights progress on the Specification for I/Os Over I3C Bridge, 14 July 2026: the two independent bridge implementations reaching I2C, SPI, UART and GPIO endpoints through one I3C controller.
  3. Logic Design Solutions — MI2CM — I2C master macro: standard mode, 7-bit and 10-bit addressing, hardware filter on SCL and SDA, no general call, system clock up to 180 MHz, VHDL source.
  4. Logic Design Solutions — MSPIS — SPI slave IP, DO254: standard and extended run-time modes, 14-bit address with 16 or 32 bit data, system clock up to 220 MHz, documentation available on request.
  5. Logic Design Solutions — MSPIM — SPI master IP, DO254: up to eight SPI slaves, programmable clock polarity and phase, serial clock period of at least six system clock periods.
  6. Logic Design Solutions — M16550 — UART based on the NS16550: NS16550 lineage, up to 1M baud, 16-byte transmit and receive FIFOs with three error bits per received byte.
  7. Logic Design Solutions — MUART — UART macro, DO254: 5 to 8 data bits, parity, break and overrun detection, register and FIFO modes, divisor of 1 to 65535 generating a 16X clock.
  8. RTCA — DO-254, Design Assurance Guidance for Airborne Electronic Hardware: the process guidance behind the DO254 designation and the A to E design assurance levels.
  9. The Linux Kernel documentation — Serial drivers: reference for the long-lived 8250/16550 driver stack that NS16550 compatibility targets.