This is absolutely wild. Rendering graphics with just combinational logic and no frame buffer is the kind of constraint that breeds creativity.
The HAKMEM sine/cosine generator is such an elegant choice - it's numerically stable in fixed-point and requires only adds and bit-shifts. Perfect for hardware. I used a similar approach once for generating test patterns in an FPGA.
The fact that you can iterate on this in simulation, then deploy to actual silicon via Tiny Tapeout for $150 is honestly mind-blowing. We're living in the future.
Good question! CORDIC and HAKMEM Item 149 are both hardware-friendly, but have different trade-offs:
CORDIC:
- Iterative algorithm (needs multiple clock cycles)
- Accuracy improves with more iterations
- Generates both magnitude and phase
- Typical hardware implementation: 12-16 iterations for decent precision
HAKMEM (Item 149):
- Single-cycle computation (just two adds per step)
- Uses the recurrence: x' = x - εy, y' = y + εx
- Accuracy depends on word width and epsilon choice
- Numerically stable in exact arithmetic if ε² < 2
CORDIC is more accurate, but takes as many iterations as you have bits of precision in your angle. Another demo called Warp in this contest used pipelined CORDIC to do atan2 on every pixel to create a tunnel, which is super impressive.
As a computer science guy who interlops in computer engineering i really want to find time to build something cool like this and tapeout. The retro architectures for rendering are simple but fun! I love the project
I recommend getting started like the author did: simulation first, then FPGA. Honestly FPGA will take you very far. I always get a kick out of being able to design my own SoC. "Hmmm I need 9 separate I2C ports... Ok, copy block, paste paste paste..." Or if you have an operation in software that's taking forever you can write an accelerator for it
I do the vast majority of my work on xilinx and it's easiest to just use the built in simulator. It's free and supports vhdl and verilog. Most support just one. For lattice and microchip work I use what the tool provides which is usually a cut down modelsim or something
The other commentator mentioned Verilator (which is indispensable in larger designs) but you may also want to grab Icarus Verilog too. It's a FOSS simulator and, unlike Verilator, is 2-bit and so it handles X ("don't care") and Z ("high impedance") signals. It's ridiculously slow compared to Verilator but the greater fidelity can be valuable depending on what you're trying to do.
Verilator is very good. It's faster than anything else, and it is free. The downsides are it won't stimulate encrypted IP blocks. And it doesn't do mixed language sim, so vhdl is no bueno.
Are there any open or at least standard FPGAs that the open source community flock to? Last time I looked into FPGAs, it was mostly closed architecture and proprietary tools
Not for anything mid to higher range, but I believe there's open source tooling for some of the older Lattice and Xilinx parts. I would say for me it's not as big a deal as on the software side, because each vendor's hardware tends to be pretty different from each other anyway.
It’s amazing and wonderful to see the Internet support these tiny cliques of interest. Having everybody connected leads to homogenization of culture in some ways, but it also supports these couple dozen (?) people around the world finding each other for this amazing little competition.
Having everybody connected leads to homogenization of culture in some ways
The internet may hypothetically homogenize culture relative to a society that does not have any kind of mass communication at all, but relative to the world it was actually introduced into, the internet has completely balkanised the culture. Prior to the internet, we had television, cinema, literature, radio, and newspapers, which were all centralised and controlled enough that they created a shared monoculture in nations. A signifant portion of a country's population would watch, read, and listen to the same media. The internet bucked that trend, allowing all kinds of new subcultures to pop up and to more easily cross national boundaries.
Yeah, back in the day you would go to school the next day after a show that everyone watches released its new episode, it aired on the prime-time slot on the primary TV channel, and you'd discuss what happened in that episode, or have some references or new jokes. Created a common culture.
I was curious about the long-term stability of the cited HAKMEM sin/cos generator. I found an overview here: https://news.ycombinator.com/item?id=3111501 (EDIT: I'm still not sure about stability, apparently it is stable in exact arithmetic under certain conditions.) Coincidentally it is related to the Verlet integration video I posted last week: https://news.ycombinator.com/item?id=46253592
Yeah, it is exact in this specific circumstance. But yes, it's exactly the same trick; I also enjoyed that video in my Youtube recommender feed last week!
> There are many bad things about LLMs, but a benign shift in popular language usage isn't one of them.
Organic shifts in language are fine. What is not fine is Big Money (which most forms of AI are) manipulating society at large - and that's not just the AI companies' doing. Think of Tiktok leading people to say "unalive" instead of the various clear words before (e.g. kill, murder, executed, run over by car, mauled to death by animal).
I disagree. It's a sign of what is essentially cultural contamination by an LLM. There is something vaguely gross about it, like when people start repeating advertising slogans. It's a sign that someone spent enough money that they directly rewired our brains.
I thought this was pretty cool but the first video didn't play. All this write up and I really just want to see the damn demo in action first! (Edit: reloaded the page and it worked. I still would like to see it on rela hardware!)
> Analog signal processing is clearly less memory than a register, no?
You are going to have a hard time doing analog signal processing with memoryless elements. In the linear domain all you can do is apply gain and mix signals together. If you work with memoryless nonlinearities you can do waveshaping, which is generally only useful when applied to special signals (e.g. sine waves).
Any time you want to do frequency-dependent behavior (filtering, oscillation) you need energy storing elements, usually capacitors, sometimes inductors. A capacitor is just like a register: it stores charge, similarly, inductors store energy in the magnetic field. Needless to say these devices are not memoryless. In fact, since the quantity that they remember is a continuous variable, they store a lot of information.
They're a kind of analogue dynamic memory. I'd hesitate to call them RAM because the Access is not Random, but they are a kind of shift register and early computers used those for RAM.
Imagine a pair of MOSFETs connected to a pair of capacitors, and a bunch of those joined together in a chain. All the gates of each one of the pair of MOSFETS are connected together, giving you a "left" and "right" clock input.
When you put a signal in if you pulse the "left" and "right" inputs, it'll store the signal voltage in one capacitor, then pass it off to the next capacitor in turn, like old-timey firefighter handing buckets of water down a line of people.
They used to use this for delaying audio signals before digital memory and analogue to digital conversion was cheap enough to use.
bucket brigades were also used to read large scale sensors like a CCD camera. they are more efficient in their use of die space because you need fewer data paths; they don't need to be digital either, each bucket can be analog for "grey" scale
>Pure Silicon Demo Coding: No CPU, No Memory, Just 4k Gates
ok, but silicon is doped so it's slightly impure, and CPUs are also silicon and memory is also silicon.
you actually meant "4K gates, no clock, no synchronization, no timing" and maybe a little "not exactly sure when the output is rea... is rea... is ready"
The HAKMEM sine/cosine generator is such an elegant choice - it's numerically stable in fixed-point and requires only adds and bit-shifts. Perfect for hardware. I used a similar approach once for generating test patterns in an FPGA.
The fact that you can iterate on this in simulation, then deploy to actual silicon via Tiny Tapeout for $150 is honestly mind-blowing. We're living in the future.
reply