F8 – an 8 bit architecture designed for C and memory efficiency [video]
fosdem.org76 points by mpweiher 4 days ago
76 points by mpweiher 4 days ago
The amount of research (or lack thereof) into whether the name was already taken in the same domain is disappointing. For me, F8 will always be this:
Question was raised (and answered) at 16:35 on the video.
The speaker’s claim that it’s hard to find a two-character name that hasn’t already been used for a CPU architecture seems ridiculous on its face. And his tone seems to indicate that he knows that (to me, anyway).
I programmed the real F8 back in the day, so I’m quite defensive about it. It’s most charming quirk: doing a long jump clobbers the accumulator.
[flagged]
Funnily enough, I did a quick search and immediately found an E8 processor - https://www.andestech.com/en/products-solutions/andescore-pr...
Wasn't quite as strong a match for "G8", but G[n] does show up in a lot of product descriptions to indicate what generation of the product is involved.
LG also put out a phone named the G8 Thinq in 2019.
I would generally agree with the speaker that it's hard not to collide when using a 2 character name. The "for a CPU architecture" narrows the collision space substantially, which does affect the full accuracy of the statement. But the spirit of "2 char IDs are collision prone" is true.
Oh, wow...never heard of the E8 or Andes, but I guess RISC-V startups are thick on the ground these days. And I can't imagine searching for G8 and not getting carpet bombed with hits for HP servers. But the fundamental question is still "why would you even want a 2 character ID in a world where searching for it (even without the massive historical name collision with F8) will make it fruitless to search for".
What a strange variation on Z80, extending it with 16-bit wide data paths... Who needs that? We have enough 16-bit ISAs around already, including PDP-11, original 8086 and even 65C816.
MSP430 has entered the chat.
I remember them selling for $4.30 when they came out
That was long after MSP430 were widely used, that pricing was part of the launch of the "value line" G-series (And it's associated MSP430 LaunchPad eval board, which is what was priced at $4.30).
Yeah I loved my LaunchPad. I bought like 20 and handed them out like candy at school
Were the LauchPads or the ST 'Discovery' line of super cheap eval boards first? Those were brilliant for getting the part in people hands.
For people comparing this to existing solutions: this is an open-source work, ready for anyone with chip production capability, with no product licensing involved or due royalties to be paid. And, given chip's simplicity, the production capability needs are not that high. Existing chip manufacturers may provide a good and convenient ready available solutions, but that's just a different thing (with you merely as a chip manufacturer's client). This is the vertical integration option, if you need it.
I wonder what's the point. 8 bits is not enough to store most values for most applications, it's bad for timers and multiplication, it's just a big waste of CPU cycles in general. The more work CPU has to do, the less time it spends sleeping, which is bad for battery-powered embedded devices. Perhaps, it has its place somewhere, but realistically, an 8-bit CPU these days is very niche at best. Imho, it's not going to take off in mainstream embedded.
Personally I'm a fan of AVR-8 which has 32 8-bit registers so if you want to do 16-bit or 24-bit or 32-bit or 40-bit math you can do it just fine, it just takes longer. The biggest AVR-8 device has about 8k of RAM and about 240k of rewritable program ROM. Those are on different buses so it can suck down an instruction at the same time it transfers data, and clocked at 16 MHz it beats the pants off any of the 8-bit micros of 1980.
You're not going to create really large systems for it, but if you want to work a gas pump or a hot water heater or make a tester for 74xx chips or a display controller for persistence of vision displays it is great.
You can code for it in C but I feel like I'm strangling puppies when I do it because it is moving the stack pointer around and doing things to support C calling conventions that I don't really need for the programs I write [1] AVR-8 assembly is fun but I still write C for it because if I need a bigger device I can recompile it for ESP32 or ARM.
Something weird about AVR-8 is that it does not have a unified address space, so in the case of that display controller, it is easy to spool graphic data out of the ROM, not so easy to upload a small amount into RAM (via serial port) and combine that with data from the ROM. I've had the crazy idea of making a AVR-8 on AVR-8 emulator (would get me to A-rank if not S-rank on AVR-8 assembly) which would make it possible to upload tiny programs into RAM but that probably requires software emulation of unified pointers for program memory.
[1] https://betterembsw.blogspot.com/2014/07/dont-overflow-stack...
I've looked at F8's ISA reference and it has lots of instructions to support 16-bit numbers, including all the basic arithmetic and bitwise operations, plus loads/stores/pushes and pops. It's almost a 16-bit ISA, actually.
Which is just bizarre since, again, we have 8086, we have MSP430. And if you are fine with most of your data being 8-bit (which is not that uncommon), there is e.g. 8051 which is still quite popular.
OTH 8-bit floating point types are so hot right now in the GPU world:
https://developer.nvidia.com/blog/nvidia-arm-and-intel-publi...
Also, a simple 8-bit CPU like the 6502 is just 3.5k transistors while simple 16-bit CPUs like the x86 or 68k are somewhere between 30k and 70k transitors (e.g. I wonder if a 6502 running at full throttle still requires less energy than an x86 or 68k doing the same work in the same time).
A minimal 16 bit system is something like a XAP [0] at 3,000 transistors. Incidentally, also a terrible architecture to program for.
I don't think any "real" x86 processor from the 286 onwards had under 100k transistors.
I haven't use these in a long time but fairly certain the memory address is more like 16 bit (otherwise, yeah how could you do anything). Still with 8 bit instructions many ops require several instructions. If that doesn't matter and it needs to be super cheap / small / etc., I think that is why these exist.
How does F8 compare to RV16 in terms of resources (die size) used and performance ?
There already is an 8bit architecture designed for C: AVR
There's also an 8-bit architecture designed for emulating 32-bit architectures: AVR!
(https://dmitry.gr/?r=05.Projects&proj=07.%20Linux%20on%208bi... ... and, yes, that's the joke, it's still an amazing hack more than a decade later!)
Oh, looks interesting, lets check out the architecture documentation! searching.... Ah, found it: Makefile and manual.tex... Yeah, f off :)
While I agree that repos with docs should have "releases", in GitHub parlances, with compiled PDFs/ebooks/whatever, building it yourself is as simple as
git clone https://github.com/f8-arch/doc.git f8-doc && cd f8-doc
docker run --rm -v "$(pwd)":/data fabtesta/ubuntu-xelatex make # or any other container with XeLaTeX that you trust
xdg-open manual.pdf
I know how to do it. The sheer audacity that I am being asked to do it rubs me the wrong way. Once I am on my phone, then I am on my work pc with no rights. Then by the time I got home i have lost my interest
Out of curiosity, what would you prefer? I wouldn't say tex is my all time favorite format, but definitely better than 90% of the documentation solutions out there.
A clickable PDF. Dont ask me to jump through loops and hoops to read about your stuff
I honestly don't see the point. I personally would prefer a CPU architecture with as wide a memory bus as possible.
All the way to 16384 bit and at least 4, but preferably 16 cores, each core with at least 1 MiB of on chip SRAM.
Now that would be useful to me.
You need 16 cores and 16MB+ of memory to control a toaster? A water heater? The washing machine? At the cheapest possible BoM and low power usage?
Because that's the kind of stuff an 8-bit micro does.
No one dreamed of controlling a toaster with an 8 bit CPU in 1969
Faggin started the 4004 in '69 and Hoff was pushing off the bat for it to be more than a calculator chip so it could be used as a general purpose processor. So they might not have specifically been 'dreaming of controlling a toaster', they and others were clearly dreaming about similar control tasks, like printers. They wrote the history down, first person...you might check it out.
Or, gee, was the comment intended to tell us how you tied an onion to your belt in 1969 (which was the style at the time) and made toast as the good lord intended without any newfangled chips in the way, up hills both ways, and we should all get off your lawn? /s
I certainly do pine for the early years of civil discourse
Oh my, pulling the "snark for me and not for thee" card. Sometimes you get the discourse you deserve.
C is undeniably a legendary programming language, but it's time to move beyond the constraints of the C abstract machine, which was heavily shaped by the PDP-11 due to Unix's origins on that architecture. C feels outdated for modern computing needs.
It lacks features like lambda calculus, closures, and coroutines—powerful and proven paradigms that are essential in modern programming languages. These limitations make it harder to fully embrace contemporary programming practices.
The dominance of C and its descendants has forced our systems to remain tied to its design, holding back progress. Intel tried to introduce hardware assisted garbage collection, which unfortunately failed miserably because C doesn't need it, and we are still having to cope with garbage collection entirely in software.
While I’m not suggesting we abandon C entirely (I still use it, like when writing C FFI for some projects), we need to explore new possibilities and move beyond C to better align with modern architectures and programming needs.
I'm pretty sure the second there is a significantly better alternative that offers the same flexibility and control, plenty of people will jump.
Same for C++.
Assuming everyone else is an idiot leads nowhere worth going.
https://ziglang.org/ is a solid future C-replacement, IMHO. There's pretty much no downsides and all upsides from a C hacker's perspective. It just hasn't reached 1.0 yet!
Newly created programming languages specifically tailored to GPUs (e.g. all the shading language dialects like MSL, GLSL, WGSL, HLSL, ...) are not limited by backward compatibility restrictions to C - and the execution model of GPUs is very different from traditional CPUs, yet all those languages turned out not much different from C extended with a handful new types and attributes.
Intels iAPX 432 had failed because it couldn't beat the much simpler, faster and cheaper, 'stop-gap' x86 design, not because of some C or PDP-11 conspiracy (and the Motorola 68k was much closer in spirit to the PDP ISA and a 'better fit' for C, yet it also lost against the rather crude x86 design).
> It lacks features like lambda calculus, closures, and coroutines—powerful and proven paradigms that are essential in modern programming languages. These limitations make it harder to fully embrace contemporary programming practices.
And what features exactly would you propose for a future CPU to have to support such language constructs? It's not like a CPU is necessarily built to "support C", since a lot of code these days is written in Java/JavaScript/Python/..., but as it turns out, roughly any sane CPU can be used as a target for a C compiler. Many extensions of current CPUs are not necessarily used by an average C compiler. Think of various audio/video/AI/vector/... extensions. Yet, all of them can be used from C code, as well as from any software designed to make use of it. If there is a useful CPU extension which benefits let's say the JVM or v8, you can be sure these VMs will use those extensions, regardless of whether or not they are useful for C.
> Intel tried to introduce hardware assisted garbage collection, which unfortunately failed miserably because C doesn't need it, and we are still having to cope with garbage collection entirely in software.
Meanwhile IBM did in fact successfully add hardware assisted GC for the JVM on their Z series mainframes. IBM can do that, since they are literally selling CPUs purely for Java workloads. With a "normal" general purpose CPU, such a "only useful for Java" GC extension would be completely useless if you plan to only run let's say JavaScript or PHP code on it. The problem with such extensions is that every language needs just so slightly different semantics for a GC and as a result it's an active research topic how to generalize this to make a general "GC assist" instruction for a CPU which is useful for many different language VMs. Right now such extensions are being prototyped for RISC-V, in case you missed it. IIRC for GC in particular, research was going in the direction of adding a generalized graph traversal extension since that's the one thing most language VMs can use somehow.
C is in no way "holding back" CPU designs, but being able to efficiently run C code on any CPU architecture which hopes to become relevant is certainly a requirement, since a lot of software today is (still) written in C (and C++), including the OS and browser you used to write your comment.
Just to be clear: this topic here is about tiny microcontrollers. The only relevant languages for such microcontrollers are C/C++/assembly. Nobody cares if it can do hardware assisted GC or if it can do closures/coroutines/... or something.