SPI: The serial peripheral interface [video]
youtube.comBen's videos are some of the most interesting content on Youtube. The build an 8-bit computer from scratch tutorial has been around this site a few times, as has the 6502 series. But what really gets interesting are these peripheral videos where he digs into a specific system or concept. His videos about how error correction works were a real eye-opener for me. It's the right level of zoomed in.
The parts when he troubleshoots are the most educational by far. He explains the theory, he puts it in practice, and debugs the result. Also, his editing is perfect IMO, he really knows how to teach.
I wish more devices simply used RxTx (or COM/RS-232) so that the protocol was defined inside the character payload instead. Simplicity is worth sacrificing bandwidth for.
A nice advantage of SPI is that it can not lock up the master.
This is not true of I2C, where the slave could stretch the clock forever.
This is also not true of UART, where the slave just might not respond. But in fact both SPI and I2C can be simpler than UART since they are strictly master/slave. With a UART, the slave could send you unrequested data, which means your software has to handle asynchronous events. Even if it doesn't you probably have to implement timeouts and input flushing.
SPI is simpler than I2C in that it is a single-master bus, so accesses are atomic (at least at the bus level, within your RTOS is another story...).
You can use simple logic gates as SPI peripherals: for example, 74HC595 works as a SPI accessible 8-bit output port. This is important because such chips are the cheapest available (vs. I2C output port or UART to parallel..)
Despite all this simplicity, SPI is very fast.
Where UART has the advantage is software: picocom on Linux or Teraterm on Windows. Basically you don't have to develop any host-side software if you use a UART. I like to use the old ymodem protocol for embedded system firmware updates.. since again no host-side software has to be developed.
UART has another advantage, fungible cheap very-long runs when paired to balanced RS485 transceivers–at least, for relatively slow baud rates, and with decent shielded cable, you can get 56 kbps well over a kilometer. Better with some effort.
In a past life I got a patent on a hardware product that used UARTs and cheap optical parts to do face-to-face wireless at 230K+... we did our firmware upgrades and configuration in the field (where installation environments might have dozens of our product).
I always liked I2C, though; and if I'd stayed in that industry, we were looking at CAN for fault tolerance...
OMG, I2C is... terrible, as others have commented, because a slave can hold up the bus 'forever'. Also, not frequently understood is the crappy protocol requires both CLK and DATA to be bi-directional in order to handle all the required states. Top that off with 3 different 'standard' speeds, two different addressing modes, non-standard voltages often requiring voltage translation -- I2C is just a terrible excuse for a bus. In fact, I've never met a practicing geek who likes I2C, as it's nothing but headaches dealing with the corner cases, and whacking the chips when the bus goes catatonic. Ugh. Just terrible.
And CAN? That's worse. I don't want to raise my blood pressure too much, so suffice it to say that CAN was invented in the 80s so it could be implemented with small amounts of code and fairly simple hardware because: 80s. It requires very careful planning of every message type. There is a limit to the # of messages you can have on any one bus. The bus is not one speed, could be many common speeds. It's a standard that needs to be put into the past ASAP.
What should we use instead? I3C is trying to supplant I2C, but SPI is clearly the simple, fast alternative. UARTs these days (e.g. the common CP2102) can do 3.8 Mbaud easily - that ain't too bad. And if you need anything approaching 'real networking' --- then, gosh, use Real Networking: Ethernet. Even poky 100 Mb/s 4-wire Ethernet is better than CAN.
(note: any technology can have a 'perfect' niche, so sure, in certain applications. I2C could be optimal, and so could CAN; but generally, I claim those are mostly legacy applications.)
I've done similar things with UARTs over long distance- I was about to write about it, but I actually have it on github:
https://github.com/jhallen/joes-sandbox/tree/master/hw/lined...
CAN is interesting in that the protocol is pretty complex, but from software point of view it's just send_packet() and register a handler for receive packets. The annoying thing is that MCUs don't have built-in CAN line drivers.
A new thing is 1000BASE-T1- Gigabit Ethernet over a single twisted pair. This is supposed to be the future in the auto world..
i2c-tools gets you pretty far if you have a userland i2c bus in linux
and spidev is not TOO hard to use if you don't need transactions.
Some FTDI chips can do I2C (also SPI), will i2c-tools work for it?
Otherwise the Arduinos with USB are a convenient way to connect I2C to a PC..
I've mostly used it with a beagleboneblack. Not sure if ftdi chips work. SPI seems hard to do with USB due to latency.
Simplicity of what exactly? From the perspective of the device/hardware, SPI is much simpler. A UART based solution tends to be relatively much more complex on the hardware/device side. You can implement a basic compliant SPI slave in a tiny bit of silicon - like a handful of flip flops simple. Moreover, you realize many of the simplest SPI devices don't have an oscillator, so you're still going to need a clock line - and if its truly a UART/asynchronous it may not be convenient to get a high enough speed external clock.
The point of SPI is that it can be implemented with a simple shift register at either end. It's not about bandwidth so much as implementation simplicity.
For SAR[1] ADC's for example, the output can be generated while the ADC conversion is happening. This reduces complexity and latency, and allows you to easily control the sampling time by how fast you clock the SPI bus[2].
If the ADC had to output as text via UART it would need to complete the conversion and then convert the output to characters before stuffing it over an UART with start and stop bits and whatnot. And you'd need some other way to control sampling and conversion time. An internal clock? External, but separate clock? Either way, added complexity.
[1]: https://www.maximintegrated.com/en/design/technical-document...
[2]: https://www.microchip.com/en-us/product/MCP3202 (semi-random example, datasheet chapter 5 and figure 5-2)
Do you mean implicit clock to save the third line? Other than that SPI is about as simple as it gets. Plus there's a protocol for practically every requirement: i2c for simple slowish shared bus with clock, 1-wire shared bus without clock, SPI fast and simple, serial single ended (RS232) or differential (RS485) for stuff outside of your PCB...
Given that I2C and SPI are addressable replacing them with a bare TTL serial interface would be horrendous because now everyone needs to invent their own addressing scheme which they already had with the other two.
0/10 suggestion
I guess with RxTx or UART as some seem to call it, you only have one device per RxTx pair so you need more IO pins... true didn't think of that, but my only real uses are 1) make some modem like thing for my C64 and 2) talk to a LoRa radio from my Raspberry Zero in both those cases I'm actually ok with having only one pheriperal and there is nothing else I want except the connection, all other things use other protocols.
SPI isn't addressable.
In my experience has been that there exist essentially 2 types of SPI, one addressable by CS, the other (more i2c-like) by the first byte (CS still required to frame the transactions.
I'm working with one of the latter type at the moment that's even worse, it daisy chains chips with each editing the address as it passes through until it hits one where it's 0, they share a CS and a wired-or interrupt, but avoid the need to have 32 CS pins
I’ve never seen SPI that doesn’t require a CS line. The whole way that daisy chaining works is this. Do you have a device off hand that uses addressable SPI, I want to see how they explain it in a datasheet.
I'm afraid the two I can think of have non-publicly available data sheets: one is a cablecrypto device, the other a SLIC. The SLIC is the one with the daisy chain - they want to be able to build phone systems with a SLIC for every slot in a T1 frame - so 24 SLICs - they're trying to avoid 24 CS pins and 24 interrupts on the attached signalling CPU
There are some Microchip chips that implement weird SPI/I2C hybrid which does I2C-like slave addressing over SPI. But thats it for the addressing.
Yes it is - that's what /SS is for.
Chip selects are optional. You can daily chain multiple spi devices and push the data out in a contiguous block. All the chip selects are tied together and only one cs line is used from the cpu.
> You can daily chain multiple spi devices
For some simple SPI devices, sometimes. Many SPI devices, especially more complex ones, won't let you shift data through the device without attempting to interpret it.
Or will shift out something completely different, like the response to the last command they received, just FFs so you can three wire in the general case, or even just garbage.
Flash memory devices use the CS line as an integral part of the protocol. It is definitely not optional.
If that counts then you can do exactly the same thing with a UART-based protocol, no?
Do you know of any uart hardware that'll only latch on /SS active?
No, I don't, but at the point where you're integrating a UART, adding a couple of logic gates on the data input doesn't seem too much of a stretch.
I'm not saying using UARTs instead of SPI is a good idea! I'm just saying the reason that it's not a good idea is not because UART's don't have /SS inputs.
But then how is the device supposed to know when it can respond? Most uart devices don't respect flow control. They assume they can transmit at any point.
Fair point.