It's fantastic to read about experiences like this. What a tremendous level of accomplishment and satisfaction they must have felt! The closest thing that I've done to this was in taking 'Nand to Tetris' which is brilliant (https://www.nand2tetris.org).
I've always wanted to play with an FPGA in this way, but I know too little to pick one that could be used to implement a CPU. They're all expensive ($50 or more) and have limits that I don't know how to interpret. For example, would this[0] be suitable for implementing a CPU soft-core? I've done some electronics before but never used an HDL and never dealt with FPGA dev boards.
I'm thinking the same. I'd like to use FPGA to create some circuits, maybe start from simple ones such as ALU and build up to CPU. Not sure if it's good idea though.
This is an impressive project, and great example of teamwork. Yeas ago I'd used uclinux on an fpga, which was designed at the time to run on simple processors without mmu support, had you looked at that option? That said it's a lot more fun for the cpu guys to build those features and see them working. Any stats on the final HW design, mhz, resources, etc?
The ability to create almost everything I need for say a laptop or a desktop would be amazing. A custom built chip that could be 3D printed maybe — maybe some derivative of RiscV will get us there sometime in my lifetime (next 50 years).
You already can get part of the way there by desigining your own FPGA board (I believe there are some Lattice chips with an open-source toolchain, otherwise the toolchain is closed). On top of that you can run your RISC-V CPU and the other components you need to create a working SoC, with all the RTL under your control. Custom silicon in our hands would be quite a stretch though.
Yes indeed - the open source toolchain (yosys, nextpnr and GHDL for VHDL support) isn't 100% "there" yet, but it is good enough to use for real projects as long as you keep the limitations in mind. It supports several different FPGA families now, with Lattice ICE40 (project icestorm) and Lattice ECP5 (project trellis) being pretty mature, and (somewhat more experimentally) project XRay for Xilinx series 7 and project Mistral for Cyclone V - and others too.
My own CPU (EightThirtyTwo) and SoC project (SDRAM, video, sound, interrupts, uart) can be built for ECP5 using Yosys and friends, as well as for a number Altera/Intel FPGAs using the proprietary Quartus.
Glad to see that open toolchain support is improving, as FPGAs on modern nodes have the potential to run truly open hardware with reasonable performance. When I did my last FPGA project I was basically forced to use the Xilinx and Altera flows from start to finish. I work in the silicon industry and it's sad that everything is proprietary (compared to the number of open-source frameworks and tools on the software side).
This is a +1 to all those that commented below/above (I never know where these will get rendered). I had no idea you could already get pretty far. That's awesome.
> CPU experiment is a little famous exercise held in the winter of the junior year in my department, the Department of Information Science at the University of Tokyo. In the experiment, students are divided into groups of four or five students. Each group designs an own CPU architecture, implements it on an FPGA, builds an OCaml subset compiler for that CPU, and then runs a given ray-tracing program on the CPU. Typically, one or two people are responsible for each of the CPU, FPU, CPU simulator and compiler.
I have seen professional programmers fail projects much easier than this. How does the University of Tokyo manage to teach their students so well, or do they show up to it better trained/motivated than the average Western student?
In order to get admitted to University of Tokyo, you have to be top on the common national university entrance examination, and then be top on the University of Tokyo entrance examination. So only the best of the best in Japan gets admitted. Highly competitive, as a degree from that University sets you for life.
I have been in a university with that much competition (except that I enrolled back when the competition was less fierce, and I observed this as a teaching assistant), and it is very frequent that the competition only lasts until the point of matriculation because university curricula are more lenient in those countries. I would not expect most CS students in the U of Tokyo would be able to do the same.
Just 'cause you're smart, it doesn't mean you know how to design CPUs (or write C Compilers.) And if the national university entrance exam is testing for VLSI design domain knowledge, they're doing something wrong.
It sounds like UTokyo is like MIT. My first test there had material that wasn't in the book and wasn't covered in the lectures. When I asked my instructor about that, he said "Oh. You're supposed to know to do research outside of class," and when I asked my class-mates they said... "Oh. Yeah. My fraternity maintains a file of past tests for each instructor. You can get a good idea for what's going to be on various tests from reviewing the files. And if you don't understand it, you can get a frat brother to tutor you on it."
(As an aside... one of the frats had THE EXACT SAME TEST I had just taken. The prof didn't even change the problems. sigh)
So... my experience with MIT was that it's a place where bright kids go to get taught by upper-classmen and the classes are there only to prove you have some sort of mastery, or at least familiarity, with the material.
At Xmas that year I was bemoaning this fact to a friend from high school. His father was a physics prof at the local state university and overheard my complaints. He offered to give me a place in the lab if I xferred over. I took him up on his offer and wound up with a desk in the undergrad office, an account on the departmental VAX (this was a big thing back in the day) and a key to the physics building and the optics lab. I could drop in to my professor's offices virtually any time and most of them were excellent in explaining the finer points of quantum chromodynamics or math methods. I was recruited to be on the "let's build a super-cheap STM" team and landed a scholarship award for my work. After taking a hiatus to defend democracy I returned and landed a part-time gig at the Superconducting Super-Collider. That year I got my own MicroVAX. (Thank you, congress for all that SSC money.) But, of course, it didn't last. (Curse you, congress for taking all that SSC money away.)
My point may be that smart kids will do well at whatever university they attend. Also, I think I did MUCH better at the local state school than I would have at MIT. (Though I didn't actually graduate w/ a Physics B.S. as I planned. IBM hired me before I graduated and it was a PAIN IN THE ** to matriculate with even a B.A.)
Also. Lori Glaze (NASA director of Planetary Science) was a class-mate of mine, so... you know... it couldn't have been THAT bad of a school.
Now... as a software development manager, the thing that REALLY impressed me was the ability of a group of kids to effectively work together. The United States pumps out a lot of very bright CS grads, and our national mythos of "the rugged individual" helps in some ways, but I'm going to guess most CS students aren't getting a lot of experience in groups larger than 2 or 3 people. When I hire recent grads from US schools, the most important thing we have to teach them is how to play well with others. Sounds like the UTokyo team already learned this lesson.
It’s about being able to get in, getting out is irrelevant. People like me are a screwed for life because we didn’t get in to any good schools in the first place.
This is not a course everyone enrolls in. Generally the top 10% who are competitive enough enroll on it. But kids have done exceptional projects in this course. The creme de la creme students of Japan are in Tokyo U or Kyoto U, so the intellectual pool is generally very good (Tokyo University generally ranks within 40 consistently in CS/EE internationally)
Thanks for the first-hand information! I was also curious about this.
> This is not a course everyone enrolls in.
This is the ticket -- at both the University of Arizona and MIT, I've seen a group of folks graduate with a CS degree after taking OS, compilers, databases, and abstract algebra. Another another group of folks graduated after taking HCI, software engineering, design, and psychology courses. The two groups had some baseline skills (all knew the basic data structures and algorithms), but otherwise appeared quite distinct.
I don't know how to phrase this formally, but I think some statement like the following is true: within-university variance is higher than between-university variance.
(When I was younger, I had strong opinions on which one of these groups were "real" computer scientists. This was a very unfortunate way of thinking that prevented me from talking to folks who I later realized were some of the smartest around. I wish someone had corrected me sooner -- solving a problem with inputs/outputs well-defined enough to apply "rigorous" techniques doesn't make those problems inherently valuable or "harder" than others. God gave all the easy problems to the physicists.)
From this side of the atlantic that always seem quite strange to me, the only way to mix so disparate set of skills is to have two degrees, the software engineering degree of the first group + the HCI and psycology stuff in a second degree for social sciences.
Thanks for sharing. Can you please share some short stories about those 10%? I'm curious what kind of education they received before they managed to pull that off. Judging from the article posted by OP, one gentleman was able to create a mini-curse "in 4 hours", which is impressive. I have a feeling that they are "Carmack-grade" programmers.
This same sort of thing was all taught on my degree course (Manchester, UK). Implement an ARM core (with a reduced instruction set), run on FPGA, write a compiler in the compilers course etc etc.
It set me up well - I've written a few emulators as a hobbyist.
There are loads of free RISC-V cores that you can read the source of and run on cheap FPGAs. Take a look at PicoRV32: https://github.com/YosysHQ/picorv32
My experience is similar - Poznan, Poland - building your own CPU architecture is part of curriculum, most students hate it (and struggle to pass or don’t pass because of it). For people who like it and get it, it’s a perfect experience.
No, it's a different issue. I did my BSc in Poland, then MSc in Germany and now finishing PhD in Switzerland. There is a huge difference in the quality of teaching caused by decades of bad practices, lack of respect towards students, conservative mindset of faculty that prevents modernization effort, and lack of motivation caused by high teaching load and poor compensation. On top of that, you get the negative selection in Polish academia that results in filling positions with mediocre and passive researchers.
Our students are not less independent or less inteligent than Germans and Swiss. It's our faculty and teaching staff that is often neither able nor willing to implement good teaching practices.
Can't speak for Poland but further east this extends to basic education too. The primary mode of education is memorization, teachers routinely insult pupils, and there is just a startling lack of understanding of how to enable a person to learn. They would much rather blame the student than look at their own methods.
I have often wondered what causes this huge difference in efficacy of teaching between eastern and western Europe. Past influence from the Soviets? Wealth difference? Other cultural influences?
I think that's related to how commenters are reacting. The OP sort of implies the only contact he had with instructors was at the beginning of the year when they received their assignment.
I left a comment above about how I left MIT (which doesn't teach as much as it provides an infrastructure for taking tests students use to demonstrate mastery) for a state school (that actually taught the courses offered and provided students with ample lab opportunities.)
I'm now thinking one of the great benefits from being taught by an actual professor (instead of an upper-classman fraternity brother) is that you get a context that is formed over the lifetime of a research career. I learned equations by reading the textbook. But how they were developed, why they were developed, what equations were used before and why they needed better models? I don't think you're going to get that from your fraternity brother.
So... yes... hopefully the UTokyo team had access to the faculty to direct them towards the most fruitful areas of the academic garden. And yes... I think there are some very good CS students out there in the states that could do the same thing. But as I mentioned above, it's great to hear a tale about smart kids working together as a team.
When instructors were well motivated and designed the course appropriately, that can surely happen. I'm not sure if this particular course is mandatory though (see my sibling comment).
Could you explain? How does not studying CS relate to being able to build a CPU/compiler/ray tracer? You don't need advanced CS to do so. Obviously the people who developed the main methods/algorithms knew their shit, academically speaking, but you can basically just engineer your way through all of this.
> How does not studying CS relate to being able to build a CPU/compiler/ray tracer?
I think it directly relates. University or not, I’m pretty sure you’ll need to study some amount of computer science to do this.
And then slightly related- When I was in college, I was able to deep dive into subjects like higher-limit just intonation and schenkerian analysis (music theory).
So to what you’re saying, I think you’re right. When the parent of my original comment said “fail”, really that’s just “giving up”, whether it be lack of time, lack of interest, or a lack of creative problem solving.
When I went to college, most universities did not offer a CS degree. (Though the university I eventually went to had a "Computer Science and Engineering" degree that mixed bits of the Math Department's Computer Science concentration and the Engineering Department's Electrical Engineering curriculum.)
I started my software career as a "that guy who can write FORTRAN" and eventually became interested in more computer-sciencey topics and less interested in particle physics. I wound up learning scheme by reading SICP and algorithm analysis by reading Knuth. I also lucked into a job where I got to talk to Ron Rivest every couple weeks, so probably got more access to him than the average MIT undergraduate.
But mostly... getting a CS degree these days teaches you what to say during interviews. For instance, I stepped a kid through linked lists when I was interviewing for a position at Amazon. He had, of course, never heard of them and confidently informed me that hash-tables were the answer to every computer science interview question (I did not go to work for that team at Amazon.)
As best I can tell, CS programs do a mix of teaching kids useful analytic skills and how to answer questions in interviews. I've hired both CS and non-CS degree holders. CS grads come with a pre-set "context" or "meta-mental-model" for how to solve computing problems. It is often very useful and I don't think you get that with other degree programs.
However... in some corners of the world, that mental model is a detriment. I've had to untrain CS grads and get them to go back to first principals in some cases.
I think the question is... is your organization doing something that would benefit from the contextual meta-model embedded in the CS curriculum at a typical university? Are you using Java? Are you programming server software on a posix-like operating system? How ambiguous are your requirements? And the EE kids get a bit more systems stuff, it seems. Half of the CS kids I've hired couldn't tell me how a larger or smaller cache would affect performance of a particular algorithm. All of the EE kids could. Managing ambiguity of requirements through judicious application of Gemma-esque design patterns seemed second nature to CS grads. The same concepts seemed to flummox EE grads.
We are all victims of our training and the best thing you can say about someone who got an art history degree and then went into programming (assuming they can code) is they learned to analyze, design and implement software w/o the benefit of a conceptual scaffolding provided for them. They COULD be the best coders for certain types of ill-defined problems; they've demonstrated their ability to construct meta-models for evaluating real-world problems (again, assuming they can code.)
> But mostly... getting a CS degree these days teaches you what to say during interviews. For instance, I stepped a kid through linked lists when I was interviewing for a position at Amazon. He had, of course, never heard of them and confidently informed me that hash-tables were the answer to every computer science interview question (I did not go to work for that team at Amazon.)
My experience of comp sci and of FAANG questions lead me to believe that this is wrong (you do learn linked lists, even if you ignore them) and an outlier (FAANG interviews often have "implement a modified linked list" style questions.
My CS course at Edinburgh had similar levels of projects, as did many others I’ve heard of - so I don’t think it’s anything particularly out of the ordinary for a good teaching university.
The thing is, “professional programming” is a very different challenge from a task like this one, and it’s hard to describe either as objectively “easier”. The provision of a clear outcome, resources, guidance, teaching time etc. can go a long way to helping deliver a good outcome - often things professional developers don’t have as much access to.
I’d bet the majority of developers I know (and respect) put on to this project would fail at it without support - similarly, an early-career student would likewise crash and burn if put on to, say, developing a front-end architecture for a modern web app. Horses for courses.
We did something similar in my engineering school in France too.
Design a simple SPARC-like CPU, make it work up to VGA/UART(and PS/2 IIRC?) I/O on an FPGA board, write a compiler for a tiny subset of Java for it, and make your OS work.
Am I right that this is all simulated in software? I mean for the "implemented on FPGA" part, is there a physical FPGA running this OS? If so, how does it talk to the rest of the computer (i.e. drivers, etc)?
> A field-programmable gate array (FPGA) is an integrated circuit designed to be configured by a customer or a designer after manufacturing – hence the term field-programmable.
> FPGAs contain an array of programmable logic blocks, and a hierarchy of reconfigurable interconnects allowing blocks to be wired together. Logic blocks can be configured to perform complex combinational functions, or act as simple logic gates like AND and XOR. In most FPGAs, logic blocks also include memory elements, which may be simple flip-flops or more complete blocks of memory.[1] Many FPGAs can be reprogrammed to implement different logic functions, allowing flexible reconfigurable computing as performed in computer software.
The above being from the FPGA Wikipedia page[0].
So yes, it is done “in software”, but FPGAs is fancy hardware that can reprogrammed to act as new hardware on the fly using software. Many are implemented as PCI devices, so they could interact with the rest of the system over the PCI bus.
Sorry, my question wasn't clear. What I am asking is whether the FPGA was being simulated in a VM, or whether the OS they built was running on a physical FPGA. And if the latter, how did the OS handle drivers for other components.
file input_file : ft open READ_MODE is input_filename;
read(input_file, c);
data <= std_logic_vector(to_unsigned(character'pos(c), 8));
And so on and so forth, to pump a file on the simulation PC into a VHDL logic array one byte at a time as a simulation of a UART.
Would be pretty funny if the above is the "wrong" repo for the story, but it is at least "an implementation" of the GAIA architecture, if not "the implementation" from the story.
The original Xv6 code uses the x86 IO ports for example[1][2][3], it would be easy to implement something similar for the soft CPU I imagine. Some good info about IO ports in the answers to this[4] SO question.
Design and build my own CPU: Check
Write my own C compiler: Check
Create a UNIX-like OS for the CPU that compiles with the C compiler: Now I have something to bookend my career with.
I used to joke that I was a "Full-Stack Engineer w/ VLSI Design Experience." But... I never wrote my own C Compiler from scratch (or an OCaml implementation.)
I'm thinking about getting nand2tetris done and go from there. None of the parts (build CPU using FPGA, bootstrapping compiler and porting an OS) seems to be particularly unreachable theoretically but of course they are all very difficult for me as a non CS guy.
