This isn't software rendering, it's called turning off spatial anti-aliasing. You can still do that in the modernest of modern games. Software rendering went out of the window because the CPU tends to have other things to do whilst playing the game besides only multiplying 4x4 matrices, which was shifted to the shader cores of GPUs. The look-and-feel that the author is discussing including shadow and texture LUTs can also be done on GPUs by passing the tables in to shaders as constant read-only storage buffers; that is not specific to software rendering. It bears mentioning that many modern software renderers (LLVMpipe/Lavapipe, OpenSWR, WARP) are basically full D3D/OpenGL/Vulkan devices in their own right, including shader compilers.
Also, modern spatial anti-aliasing works by multisampling within a pixel (or even neighbouring pixels) and taking the average, rather than filtering across singly-sampled pixels, which massively reduces the quality. Not to mention temporal and machine learning AA techniques including DLSS, XeSS, FSR, and all the TAA implementations in many games.
Real players know you us the coarsest mipmap level the whole time, so every texture is something like 4 pixels. And 320x200 and directmouse. It's the smooth movement, high fps and direct control that provides the sense of "being there". Not the graphical fidelity.
How is it not software rendering? These games shipped without hardware renderers. Yes, they also don't have AA, but how does that make it not software rendering?
Here’s an analogy. Suppose someone said, “I love manual transmission cars” and then described in detail the physical knobs and dials that all the cars from the 1950s had.
It turns out what they like is physical knobs and dials. You can have those in automatic transmission cars. It so happens that most modern automatics also have shitty touchscreens.
(I know this analogy isn’t precise because there was a long period of automatics with physical controls, but bear with me.)
The point is that there’s nothing _inherent_ to software rendering that’s responsible for the thing he likes. The distinction is useful because, for example, he could go seek out modern games with the look he likes. Or make games. And if he thought the thing he likes comes from having a software renderer, he’d leave a lot of performance and power efficiency on the table for no good reason. He likes a certain rendering style that’s equally achievable with both GPU and CPU rendering. The choice of hardware should probably be driven by considerations like performance, debuggability, platform support, availability of libraries, etc.
"Nothing to do" is a bit strong. It has something to do, and the explanation is in the history of how these games worked.
Can it be simulated on GPUs? Sure. But it was typical of software renderers of that era, and NOT typical of GPUs, and it has to do with how they were used.
This is incorrect. It is exactly the same algorithm as circuits in the GPU or as instructions on the CPU. No difference, except one: the way data is pipelined on the GPU makes texture blending free, or nearly so. So to show off, every 3D game back then turned it on (an OpenGL/Glide/D3D configuration option) and everything got smoothed out, often with no option to turn it off. That's why it ends up looking different: because the game itself selected different interpolation in accelerated vs software rendered. It was a game choice.
These days you can use different presets to override those choices, and have your 8K, 240Hz Quake look software rasterized, if you want.
"Software" isn't the point. You can do the exact same math on CPU or GPU. OpenGL supports nearest neighbor texture filtering on whatever hardware you run it on since 1992, one year before Doom even came out.
In a software-rendered game, the CPU draws every pixel. There’s no 3D card involved: the game works out each frame in its own code and writes it straight to video memory
I think that’s the standard definition. If you mean that this style can be achieved with graphics hardware then that’s a different argument and why your first sentence is deemed contentious
What they meant was: "This aesthetic isn't exclusive to software rendering". I thought the rest of their comment made that pretty clear once you got past the first sentence.
> If you mean that this style can be achieved with graphics hardware
I really ought to have been more specific with my first sentence to ward off any hair-splitting.
The author claimed to love software rendering, but the examples given to highlight the supposed positives of software rendering can all be achieved on GPUs. If the author had instead highlighted the ability to easily debug and inspect the inner workings of the rasterisation pipeline with a software renderer, that might have been a much more relevant argument.
Not necessarily. Someone unfamiliar with graphics might believe that there’s something about a CPU that a graphics card doesn’t have such that the look can’t be achieved. Certainly the author of this piece seems to believe that.
He likes the look of nearest neighbor texture sampling, not software rendering.
You can do nearest neighbor in hardware rendering and you can do bilinear in software rendering. Rendering mode and texture sampling are two totally different things.
But if you want to experience GLQuake that looks like software-rendered Quake, that's a simple settings change in your GPU driver, overriding the texture sampling mode. That's literally the only difference. I'm not sure the author is aware that it is a viable option to tweak the hardware rendering this way.
May be a couple settings. IIRC full bright textures didn't translate well in early GPU renderers. Often they'd appear shaded as if they weren't light sources at all.
Are you sure? I haven't played enough Quake to know either way, but I did play Deus Ex in software mode (unreal engine), and that had a lot of differences to the other renderers. I very much doubt texture sampling mode alone would reach a similar result.
(It even had graphic errors, green textures in the faces etc giving everything a very alien look. But that's not what I mean).
TAA has a pretty bad reputation among many gamers for being too blurry. TSR (temporal super resolution) is a newer implementation that's overall sharper, better at eliminating "jaggies", and performant, without visible ghosting or artifacting.
Indeed. Personally I am quite old-school, I like MSAA (despite the performance hit) and subpixel morphological AA (SMAA, used to great effect in Crysis). DLSS is interesting but I cannot unsee the artifacts.
TAA is a nightmare, and Unreal Engine depends way too much on it. TSR also suffer from ghosting and artifacts unfortunately.
But I have come to the realisation that gaming will be about massive compromises for the next decade.
"Software rendering went out of the window because the CPU tends to have other things to do besides just multiply 4x4 matrices, which was shifted to the shader cores of GPUs."
Hardware rendering took over because generating a 3D scene is embarrassingly parallel, and the industry quickly learned that you can have thousands of reasonably simple "cores" that can divide and conquer a scene in parallel. There was a no way a general purpose CPU could keep up.
Not so quickly. The shader cores were a breakthrough from early 2000s. First consumer GPUs were built as same as the 1980s GPUs but everything had been integrated in a single chip.
It is the sheer amount of compute and memory bandwidth that mogs a CPU but the parallel processing doesn't help much, at least in realistic 3D: an 8 core AVX-512 march CPU has an order of magnitude more SIMD lanes than something like RSX in PS3 yet it still wouldn't be able to match its performance in rendering textured z-buffered triangles. 3D graphics can really go wide only on wireframe/points, z-buffer forces sequencing between polys so the parallelism there is very limited.
Parallel processing was fully in force in the very first GPUs, like the Voodoo. Not like a modern GPU with thousands of more flexible shader cores, but the first iteration had a pipeline architecture that saw multiple blocks operating on rendering separate parts of the pipeline in parallel. This was revolutionary. Especially given that CPUs at the time had a single core.
I'm not sure what the PS3 comparison is for, or why we're comparing a 2026 AVX-512 CPU with a 2006 gaming console. Okay? A modern GPU can obliterate any CPU in SIMD, presuming you conform with the types it wants. Add that it's weird to bring up as AVX and other SIMD techniques are basically the same sort of architecture, parallelizing a problem.
> 3D graphics can really go wide only on wireframe/points, z-buffer forces sequencing between polys so the parallelism there is very limited.
Rendering a 3D is scene is the very definition of embarrassingly parallel. While optimizers can cull the scene, you know that the entire depth within the frustum (there are a variety of optimizing techniques to minimize the overdraw, but it's a part of rendering a scene) is "rendered" and in amalgamation the Z-buffer values are compared and discarded, blended, or whatever? It is an extraordinarily parallel operation.
I'm going to guess if you did it on the video card and simply disabled those enhancements it would be considerably more energy efficient than doing it on the CPU as well which (a) means your laptop battery lasts longer on a charge, and (b) also opens up the possibility of not needing a particularly powerful CPU in order to juice the resolution, because the video hardware is doing the work for you.
From a programming perspective I do completely get the desire to have the CPU do all the work though - if you want absolutely everything under your control it's easier, and especially if your game is single-threaded.
I've mentioned this elsewhere, but nowadays, anything you see out of an HDMI/DP/eDP cable comes out of a GPU. Modern OSs make it impossible to do 'software rendering' any more; it's just that much faster and that much more efficient, even if it is harder to program directly. Even IPMI BMCs have a tiny GPU doing stuff.
Why is it “impossible”?
If you mean that a rendered frame has to reside in the graphics memory before being sent to the display, then I think that was the case even when software rendering was a thing.
If you mean rendering a frame entirely in the main memory by the CPU, then there is nothing making it impossible. I mean, most major 3D rendering software still have CPU renderers as an option. Slow, but doable.
P.S: I am using offline renderers here as an extreme example. A game renderer - especially for scenes like Quake’s - are definitely going to be far simpler and cheaper.
By 'impossible', I mean it in a very strict sense of blitting full-screen frames directly into the framebuffers provided by HDMI or (e)DP representing a visible surface. This is still kind of possible in Windows when games are properly full screen and not 'borderless full screen', but the OS still does compositing in the background, and the GPU is still involved to some extent.
Of course, offline rendering is a different story.
It's also because at this point the operating systems themselves use the accelerated graphics architecture for things like window compositing, because that also turned out to be faster. And it's much easier to do cute little visual effects like wiggles and window distorts if those entities are actually 3D rendered graphics objects.
Even the standard desktop experience is a 3D graphics application these days.
Concur. Feels like a conflation; CPU vs GPU rendering doesn't change the appearance directly. The article performs further conflations later re the OS comparison and Steam re Ion Fury etc.
It just feels wrong historically. When HL came out, people were very enthusiastic about texture filtering and the game was clearly designed with hardware rendering in mind.
I like when games use nearest, or dedicated pixel art shaders. Bilinear just looks low-effort to me, especially when it's used for whole-screen scaling! Apple's Lanczos also looks a bit disgusting to me, I can't tolerate the display scaling on iPad because of this.
You know what I found this past year was that I could NOT find any way, to turn off anti-aliasing, for text inside a webview. I made a whole entire custom terminal app just for myself in Rust, Tauri, xterm.js and spent a ton of time trying to un-blur the text but failed. I redid the entire thing (well, Codex/Claude mostly) using a rust graphics library (egui) and now the text is finally sharp! and its fast. I am growing more and more anti webview with these AI coding tools.. they can almost one-shot convert a slow webview app into something that uses 1/10th the ram and 10x as fast. The text looks soooo much better, all fonts look better. In xterm.js, text looks super fat all the time (to me).
In the first waves of multimedia PCs - the CPU was tasked with doing sound, networking, hard drive control, etc. - pretty much all of the cost savings came from having minimal hardware for the auxiliary components and doing it all in drivers.
A common "Trick" was to buy cards that had essentially their own processors to handle this stuff - SoundBlasters, 3Com nics, IDE drive controllers that had their own processor (hell, even up to their own OS) since that freed the CPU up for doing actual work - gaming. Even an external or other
CPUs were a bottle neck but most people didnt tax the single core machines doing most daily work (office work, browsing the web) so it was fine, but once you load up a game that needs sound, 3d, networking, disk access all at once it would start stealing clock cycles trying to do all of that in one place.
I think it's because adding a sound card offloaded sound processing from the CPU and it had more cycles to draw frames! Probably back when we were at like 25Mhz
Off-loads sample mixing to the sound card. Computers of that era were slow enough that doing 44100 additions per second was expensive (IIRC quake had 8Khz samples anyway).
The sound card took sound computation off the CPU as well. Back in the day, every aspect of the game was fighting for CPU cycles.
Way back in the day there were some very impressive (given the technology) arcade ports to things like the Apple ][. That system's audio required running an assembly function that literally touched a memory address over and over again at a specific time rate; each touch flipped the electromagnet on the speaker. So games with theme music had to time slice hanging out in the audio routine versus doing anything else.
I grew up with a copy of Moon Patrol on Apple ][c and I didn't know for years that the theme song isn't supposed to sound like it's being played through a tea kettle.
CPUs from that era were single core and pretty slow. They had to do everything like game logic, graphics, sound, etc all at once, so it was often resource starved. Sound took some resources, sometimes significant depending on the sound and what else the cpu was doing, unless it was a really simple sound (think NES style sound). Moving that to a dedicated sound card meant the cpu didn't have to worry about sound anymore.
Sound, like a lot of things, could be seen as a peripheral and a lot of peripherals back then needed their own cpu to perform well. One interesting fact I recently learned is the commodore 64 was so resource starved, the external floppy drive you could buy had its own 6502, just like the C64 did as its main processor. It just had to off-load that to another cpu because its own cpu was too busy to reliably run a disk drive. There are other reasons for this like commodore's experience in drives previously was this kind of architecture so they went with what they knew, but its a bit interesting to know your floppy drive was essentially another C64 with its own commodore OS and ram and 6502.
Software rendering often provided better visual effects than the GPU paths could provide at the time. A general purpose CPU can technically do any effect since the software is the pixel shader.
Programmable shader pipelines weren't a thing until Geforce 3 in 2001.
I recently got to try out Quake 2 on 9front, a Plan 9 fork, which ran surprisingly fast on my very, very modest Thinkpad T410i with a 2.4 GHz dual core i3 at 1440x900. Plan 9 has no GPU support so all rendering is software. Even had a little lan party with other devs. I really enjoyed the software rendering, it bought back early memories of playing Q2 before we bought our Voodoo2. I even used my ploopy trackball to play. Fun times.
Presently, Plan 9 in general lacks 3D graphics but has an in-kernel 2D rendering engine referred to as devdraw. Quake 2 software rendering is done in-engine by rendering full frames to a bitmap displayed by devdraw. A fun experiment was to rcpu into my industrial DIN rail PC, a 1.2 GHz 4 core Atom, from the Thinkpad and play Quake 2. When I run Quake 2, it runs on the CPU server but graphics are sent to my laptop. However, since Q2 is rendered in-engine on the remote iPC, full frames are copied over gigE to my laptop. Q2 full screen was stuttering badly but at about 1/4 screen size it was surprisingly playable.
An other example is a port of cave story done by a 9dront dev where he integrated devdraw so the game renders by loading its assets into devdraw and issuing rendering commands. It is likely the first Plan 9 native game. It plays perfectly over an rcpu session at full screen as the rendering is done by the client side devdraw.
Another 9front dev is doing a lot of graphics work and implemented affine warping in devdraw which is useful for rapidly transforming bitmaps, e.g. scaling the output of the gameboy emulator and Doom. He and others are interested in implementing a concurrent 3D rendering software devdraw engine as we have many-core machines now. When GPU's first hit the streets, we were on single core machines. SMP machines were not affordable by mere mortals when the 3Dfx Voodoo hit the market. So the idea is, why struggle with the gargantuan task of implementing a "modern" GPU stack when you have all these CPU cores now? Why not take a step back, think about what we have learned, and extract all the performance we can from modern many-core CPUs with all sorts of vector extensions. In a Plan 9 context, you could run the game on a many core machine to handle game logic and play remotely with a bigger many-core spraying pixels at a display. Distributed gaming. Might even be able to do multiplayer with a single engine instance with multiple rendering clients attached. Fun system to play with.
I know this is not entirely what this article is about, but I love software rendering for desktop apps. It seems to me that the performance impact with proper paint invalidation is negligible, but doesn't suffer from the same driver issues as hardware accelerated rendering seems to have. Bonus that it doesnt use any GPU resources in the age of AI
The article says that "even a 3.3GHz quad-core Xeon “isn’t quite enough to run Shadow Warrior at full 1600 by 1200”." But we are on DOS, so I guess three cores are just not doing anything there?
That quote about Shadow Warrior doesn't pass the smell test. You're right there must be bottlenecks in the way.
The Build engine's rendering was stunningly simple (and dare I say: elegant). You could run hundreds of instances of Shadow Warrior on a single EPYC (since they mentioned Xeon) at that resolution and framerate if you rewrote the engine with modern mechanical sympathy in mind. On a modern desktop CPU it'd probably be over a hundred instances.
People, including many programmers, just don't grasp how much power we have under the hood nowadays...
Edit: my old iPhone SE, fully utilized, can produce the same GFLOPS as the Cray T3E (the one with thousands of DEC Alphas inside). This is our reality.
Quake DOS is from a year earlier to Shadow Warrior and it would get around 60fps at 320x200 on a Pentium II 266. The Xeon from that video is from 2008, and from geekbench 2 results it's ~30 times faster than the pentium. 1600x1200 is 30 times the pixels of 320x200.
Additionally Shadow Warrior used some newer rendering techniques, so it's possible they don't scale well to higher resolutions.
Software rendering was useful for me for a long time but as a long-time gamer (from the original DOOM to Cyberpunk 2077 and beyond) I can firmly say that real-time raytraced games at 4K/60FPS with dynamic lighting, reflections, massive geometry detail, etc, all greatly enhance the experience. That said, the old DOOM aesthetic pre-Voodoo was definitely nice. The voodoo smoothing had a nice cheap feel to it, too.
The difference in performance is just so big, unless you're targeting something without a GPU, there is no world in which you would want to leave it unutilized IMO, even if you're just rendering an UI. I can do things on my slow phone in WebGL that would stutter on my desktop if I drew the pixels with the CPU.
I share the same opinion as the author. That said, it's also important to remember that the display devices of the time were different than they are now and the sharp pixelated textures we see now would look a bit different on a CRT from that period. The sharp, gritty look that he (and I) admire would have been just a little fuzzier back then.
A tiny bit fuzzier but not in a way that mattered back then. If you're running a software renderer it could well have been at 320x200 and on any monitor that comes out still looking pretty much like an LCD display, with slight blending on the edges technically but still large, large pixels not being blended. Even at 640x480 there was noticeable blockiness not being blended away. Monitors were designed to be sharper than the televisions the consoles were outputting to. Software rendering at anything high enough to be blended looking was a slideshow.
> Ion Fury, a Build engine game from 2019, still ships with a software renderer.
Amazing game BTW, best shooter I've played in many years. I never really got into build engine games back then, so it's not just my rose-tinted glasses, that game is a masterpiece.
Ion Fury uses a modified version of eDuke32 which can do software only rendering, and quite well, but defaults to GPU acceleration because Ion Fury uses the real time lighting.
That said, the Build Engine had software voxels in 1997. Which was pretty cool.
A few years ago I seriously contemplated making a game using software rendering, in part because the only target was SDL. I actually got a 3D engine running 2K@60Hz with HL2-level scenes, with the CPU basically idle, so it's not a terrible idea.
(Then LLMs started taking off and I decided I needed to reskill out of the industry, just in case)
Unreal (and Unreal Tournament), Thief, Quake II (a slight improvement over Quake), Half-Life (a great improvement over Quake), Shadow Warrior,
Any modern port of Doom supporting software rendering, Heretic, Hexen, Descent, System Shock, Homeworld.
Quake II is a completely different game from Quake, isn't it? The levels, artwork and gameplay feel quite different. I personally prefer the look and feel of the original Quake, particularly when using the Arcane Dimensions mod.
Software renderer of Quake II has no colored lighting. It's OpenGL-only feature. Half-Life and Unreal do this better, they support colored lighting in their software mode.
I love Software Rendering too! I'd add Unreal Tournament to the list of "Games worth trying", though not sure what the multiplayer scene is like nowadays.
I am still a big fan of CPU render for VFX (not games). It is stable, capable of very large scenes. GPU is fast but you will never render something really big. And CPU look is much much better.
> you will never render something really big. And CPU look is much much better.
Why do you believe these things? In VFX and film, the GPU shaders and rendering is identical to CPU, they use the exact same programmable shaders on either side, so there is no difference in look or capability. Today’s GPUs have out of core streaming and caching and access to host memory. Pixar showed off their pixel-matching between GPU and CPU at Siggraph this year, and mentioned that today’s consumer GPUs are big enough to render IIRC ~98% of their scenes without much ‘out-of-core’ tech. Lots of film and VFX studios are using GPUs in production today to render pretty big shots. I think your story is a bit out of date; GPUs being size limited was somewhat true in the past, but is quickly going away. The CPU vs GPU look question has never been true for VFX or film, and was only true in games if you chose different rendering algorithms.
I don’t know much about this so I’m curious. Is the reason why GPUs aren’t preferable for large scenes due to their limited VRAM or a different aspect of them?
Its so interesting to me that back then we would have died for the graphics we have today, and now this retro look is being appreciated. I mean, it definitely has an aesthetic but its generally unappealing. I guess nothing is immune to nostalgia and pining for being at the age when we had less responsibilities and life seemed easier.
I don't think it really applies when we're talking about PC gaming. Consoles were blurry as hell on a huge interlaced low-res TV, but late-'90s PC games on a decent monitor had nice and crisp pixels on a CRT that weighed as much as a large dog. Especially when flat-glass trinitron CRTs and the like came out (yes spellcheck, that's a word), I don't think the end-user appearance on the screen was all that different from what we have now, other than obvious stuff like aspect ratio and incremental improvements to resolution and framerate. Maybe deeper blacks and lower latency on CRT.
The ability to shift resolutions natively was neat, though, I missed that in the pre-4K days.
Because the pipelines lacked fast, dedicated linear algebra engines, you saw a lot of clever hacks to get around having to do full linear algebra as well.
The original Marathon allowed you to look up and down, but it wasn't actually tilting the camera frame; it was applying a distortion on the field of view and then basically vertical scrolling the render target. Essentially, you can think of it like you were playing the game in vertical phone mode all the time but you were looking at the middle third of that possible output domain, and looking up and down pretty much just scrolled what you were looking at and applied a trapezoid transform.
This is why you couldn't look straight up or straight down, the trapezoid transform would become painfully obvious quickly and the actual algorithm did not really tilt your field of view.
But it was good enough that they could slap it on the box as a novel feature for Wolfenstein style first person shooters.
I remember when I finally saw my favourite 3D games (Quake 2 and Pod) on a Voodoo2 and I immediately bristled: I hated the blurry way the textures looked with bilinear filtering, and was iffy on their love of yellow lighting.
I've softened on the issue of non-white lights, but I 100% am still full of nostalgia about unflitered low-res textures.
“Software rendering” usually means “doing rasterization on the CPU”, as opposed to “hardware rendering” that is done by the rasterization primitives of the GPU.
Nanite is a bit weird because it doesn't use the hardware rasterization of the GPU, but it's still doing rasterization on the GPU, so it doesn't really fit this dichotomy.
No software rendering requires DOS, you could use it with Windows back in the day (how I originally played Half Life) and engines that still support it will run on modern OSes too.
Eh, Nanite rasterisation is still happening on the GPU, it's "just" implemented in shaders rather than using the fixed-function pipeline. "Software rendering" is becoming less useful as a descriptor as graphics hardware becomes more general.
Wrote a little software rasterizer for a roguelike years back and the best part was just being able to step through the whole pipeline in a debugger. No driver weirdness, no shader compile quirks, the pixel is wrong because your math is wrong.
Yeah, that is the hardest part about 3D graphics programming, at least it was the last time I had my elbows deep in it.
You're not doing traditional application programming once you get into 3D graphics. You are driving a microcomputer at arm's length. Every aspect of that challenge comes into play, from the fact that it's a black box to how you can have it working perfectly in one configuration of hardware, and then find it doesn't work on another configuration.
My favorite reasons for this include "there is a legitimate bug in the closed-source shader compiler" and "this graphics card advertises a capability but in reality it implements it in software so if you turned that flag on your render speed suddenly went to seconds per frame."
The CUDA era helped some with this problem because there was suddenly massive pressure on making the GPU into less of a bundle of pipelines and specialty hardware accelerators strapped together and more into standardized massively parallel fabric, because card manufacturers suddenly found themselves not wanting to be left out in the cold if they invested their time in building a very very fancy specialty hardware circuit for doing texture manipulation or something... And that proved to be useless for mining Bitcoin, or whatever they suspected might replace Bitcoin in the future. I think there are far less gotchas now regarding weird architecture than there were in the 2010s.
Also, modern spatial anti-aliasing works by multisampling within a pixel (or even neighbouring pixels) and taking the average, rather than filtering across singly-sampled pixels, which massively reduces the quality. Not to mention temporal and machine learning AA techniques including DLSS, XeSS, FSR, and all the TAA implementations in many games.
It turns out what they like is physical knobs and dials. You can have those in automatic transmission cars. It so happens that most modern automatics also have shitty touchscreens.
(I know this analogy isn’t precise because there was a long period of automatics with physical controls, but bear with me.)
The point is that there’s nothing _inherent_ to software rendering that’s responsible for the thing he likes. The distinction is useful because, for example, he could go seek out modern games with the look he likes. Or make games. And if he thought the thing he likes comes from having a software renderer, he’d leave a lot of performance and power efficiency on the table for no good reason. He likes a certain rendering style that’s equally achievable with both GPU and CPU rendering. The choice of hardware should probably be driven by considerations like performance, debuggability, platform support, availability of libraries, etc.
Can it be simulated on GPUs? Sure. But it was typical of software renderers of that era, and NOT typical of GPUs, and it has to do with how they were used.
These days you can use different presets to override those choices, and have your 8K, 240Hz Quake look software rasterized, if you want.
I really ought to have been more specific with my first sentence to ward off any hair-splitting.
The author claimed to love software rendering, but the examples given to highlight the supposed positives of software rendering can all be achieved on GPUs. If the author had instead highlighted the ability to easily debug and inspect the inner workings of the rasterisation pipeline with a software renderer, that might have been a much more relevant argument.
Any look in older games, 2d or 3d, can be achieved on modern GPUs, so that seems like a vacuous thing to state.
You can do nearest neighbor in hardware rendering and you can do bilinear in software rendering. Rendering mode and texture sampling are two totally different things.
(It even had graphic errors, green textures in the faces etc giving everything a very alien look. But that's not what I mean).
aka (just) anti-aliasing, we are talking original quake 1 era, after all.
Hardware rendering took over because generating a 3D scene is embarrassingly parallel, and the industry quickly learned that you can have thousands of reasonably simple "cores" that can divide and conquer a scene in parallel. There was a no way a general purpose CPU could keep up.
It is the sheer amount of compute and memory bandwidth that mogs a CPU but the parallel processing doesn't help much, at least in realistic 3D: an 8 core AVX-512 march CPU has an order of magnitude more SIMD lanes than something like RSX in PS3 yet it still wouldn't be able to match its performance in rendering textured z-buffered triangles. 3D graphics can really go wide only on wireframe/points, z-buffer forces sequencing between polys so the parallelism there is very limited.
I'm not sure what the PS3 comparison is for, or why we're comparing a 2026 AVX-512 CPU with a 2006 gaming console. Okay? A modern GPU can obliterate any CPU in SIMD, presuming you conform with the types it wants. Add that it's weird to bring up as AVX and other SIMD techniques are basically the same sort of architecture, parallelizing a problem.
> 3D graphics can really go wide only on wireframe/points, z-buffer forces sequencing between polys so the parallelism there is very limited.
Rendering a 3D is scene is the very definition of embarrassingly parallel. While optimizers can cull the scene, you know that the entire depth within the frustum (there are a variety of optimizing techniques to minimize the overdraw, but it's a part of rendering a scene) is "rendered" and in amalgamation the Z-buffer values are compared and discarded, blended, or whatever? It is an extraordinarily parallel operation.
> "3D cards smoothed textures with bilinear filtering"
You can just not do that.
From a programming perspective I do completely get the desire to have the CPU do all the work though - if you want absolutely everything under your control it's easier, and especially if your game is single-threaded.
If you mean rendering a frame entirely in the main memory by the CPU, then there is nothing making it impossible. I mean, most major 3D rendering software still have CPU renderers as an option. Slow, but doable.
P.S: I am using offline renderers here as an extreme example. A game renderer - especially for scenes like Quake’s - are definitely going to be far simpler and cheaper.
Of course, offline rendering is a different story.
Even the standard desktop experience is a 3D graphics application these days.
[1]: https://news.ycombinator.com/item?id=44966266
A common "Trick" was to buy cards that had essentially their own processors to handle this stuff - SoundBlasters, 3Com nics, IDE drive controllers that had their own processor (hell, even up to their own OS) since that freed the CPU up for doing actual work - gaming. Even an external or other
CPUs were a bottle neck but most people didnt tax the single core machines doing most daily work (office work, browsing the web) so it was fine, but once you load up a game that needs sound, 3d, networking, disk access all at once it would start stealing clock cycles trying to do all of that in one place.
Way back in the day there were some very impressive (given the technology) arcade ports to things like the Apple ][. That system's audio required running an assembly function that literally touched a memory address over and over again at a specific time rate; each touch flipped the electromagnet on the speaker. So games with theme music had to time slice hanging out in the audio routine versus doing anything else.
I grew up with a copy of Moon Patrol on Apple ][c and I didn't know for years that the theme song isn't supposed to sound like it's being played through a tea kettle.
Sound, like a lot of things, could be seen as a peripheral and a lot of peripherals back then needed their own cpu to perform well. One interesting fact I recently learned is the commodore 64 was so resource starved, the external floppy drive you could buy had its own 6502, just like the C64 did as its main processor. It just had to off-load that to another cpu because its own cpu was too busy to reliably run a disk drive. There are other reasons for this like commodore's experience in drives previously was this kind of architecture so they went with what they knew, but its a bit interesting to know your floppy drive was essentially another C64 with its own commodore OS and ram and 6502.
Programmable shader pipelines weren't a thing until Geforce 3 in 2001.
Presently, Plan 9 in general lacks 3D graphics but has an in-kernel 2D rendering engine referred to as devdraw. Quake 2 software rendering is done in-engine by rendering full frames to a bitmap displayed by devdraw. A fun experiment was to rcpu into my industrial DIN rail PC, a 1.2 GHz 4 core Atom, from the Thinkpad and play Quake 2. When I run Quake 2, it runs on the CPU server but graphics are sent to my laptop. However, since Q2 is rendered in-engine on the remote iPC, full frames are copied over gigE to my laptop. Q2 full screen was stuttering badly but at about 1/4 screen size it was surprisingly playable.
An other example is a port of cave story done by a 9dront dev where he integrated devdraw so the game renders by loading its assets into devdraw and issuing rendering commands. It is likely the first Plan 9 native game. It plays perfectly over an rcpu session at full screen as the rendering is done by the client side devdraw.
Another 9front dev is doing a lot of graphics work and implemented affine warping in devdraw which is useful for rapidly transforming bitmaps, e.g. scaling the output of the gameboy emulator and Doom. He and others are interested in implementing a concurrent 3D rendering software devdraw engine as we have many-core machines now. When GPU's first hit the streets, we were on single core machines. SMP machines were not affordable by mere mortals when the 3Dfx Voodoo hit the market. So the idea is, why struggle with the gargantuan task of implementing a "modern" GPU stack when you have all these CPU cores now? Why not take a step back, think about what we have learned, and extract all the performance we can from modern many-core CPUs with all sorts of vector extensions. In a Plan 9 context, you could run the game on a many core machine to handle game logic and play remotely with a bigger many-core spraying pixels at a display. Distributed gaming. Might even be able to do multiplayer with a single engine instance with multiple rendering clients attached. Fun system to play with.
The Build engine's rendering was stunningly simple (and dare I say: elegant). You could run hundreds of instances of Shadow Warrior on a single EPYC (since they mentioned Xeon) at that resolution and framerate if you rewrote the engine with modern mechanical sympathy in mind. On a modern desktop CPU it'd probably be over a hundred instances.
People, including many programmers, just don't grasp how much power we have under the hood nowadays...
Edit: my old iPhone SE, fully utilized, can produce the same GFLOPS as the Cray T3E (the one with thousands of DEC Alphas inside). This is our reality.
Additionally Shadow Warrior used some newer rendering techniques, so it's possible they don't scale well to higher resolutions.
Sources: * https://framebuffer.io/project/quake * https://browser.geekbench.com/geekbench2/327940 * https://browser.geekbench.com/geekbench2/529270
Amazing game BTW, best shooter I've played in many years. I never really got into build engine games back then, so it's not just my rose-tinted glasses, that game is a masterpiece.
That said, the Build Engine had software voxels in 1997. Which was pretty cool.
(Then LLMs started taking off and I decided I needed to reskill out of the industry, just in case)
I can add more:
Unreal (and Unreal Tournament), Thief, Quake II (a slight improvement over Quake), Half-Life (a great improvement over Quake), Shadow Warrior, Any modern port of Doom supporting software rendering, Heretic, Hexen, Descent, System Shock, Homeworld.
That colored lighting, though.
Why do you believe these things? In VFX and film, the GPU shaders and rendering is identical to CPU, they use the exact same programmable shaders on either side, so there is no difference in look or capability. Today’s GPUs have out of core streaming and caching and access to host memory. Pixar showed off their pixel-matching between GPU and CPU at Siggraph this year, and mentioned that today’s consumer GPUs are big enough to render IIRC ~98% of their scenes without much ‘out-of-core’ tech. Lots of film and VFX studios are using GPUs in production today to render pretty big shots. I think your story is a bit out of date; GPUs being size limited was somewhat true in the past, but is quickly going away. The CPU vs GPU look question has never been true for VFX or film, and was only true in games if you chose different rendering algorithms.
Now that will truly blow his mind, at least on this setup and older games.
The ability to shift resolutions natively was neat, though, I missed that in the pre-4K days.
The original Marathon allowed you to look up and down, but it wasn't actually tilting the camera frame; it was applying a distortion on the field of view and then basically vertical scrolling the render target. Essentially, you can think of it like you were playing the game in vertical phone mode all the time but you were looking at the middle third of that possible output domain, and looking up and down pretty much just scrolled what you were looking at and applied a trapezoid transform.
This is why you couldn't look straight up or straight down, the trapezoid transform would become painfully obvious quickly and the actual algorithm did not really tilt your field of view.
But it was good enough that they could slap it on the box as a novel feature for Wolfenstein style first person shooters.
(And I actually do like their sound, but I wouldn't want to be stuck with only tubes)
I've softened on the issue of non-white lights, but I 100% am still full of nostalgia about unflitered low-res textures.
Nanite is a bit weird because it doesn't use the hardware rasterization of the GPU, but it's still doing rasterization on the GPU, so it doesn't really fit this dichotomy.
A mix of software and gpu
You're not doing traditional application programming once you get into 3D graphics. You are driving a microcomputer at arm's length. Every aspect of that challenge comes into play, from the fact that it's a black box to how you can have it working perfectly in one configuration of hardware, and then find it doesn't work on another configuration.
My favorite reasons for this include "there is a legitimate bug in the closed-source shader compiler" and "this graphics card advertises a capability but in reality it implements it in software so if you turned that flag on your render speed suddenly went to seconds per frame."
The CUDA era helped some with this problem because there was suddenly massive pressure on making the GPU into less of a bundle of pipelines and specialty hardware accelerators strapped together and more into standardized massively parallel fabric, because card manufacturers suddenly found themselves not wanting to be left out in the cold if they invested their time in building a very very fancy specialty hardware circuit for doing texture manipulation or something... And that proved to be useless for mining Bitcoin, or whatever they suspected might replace Bitcoin in the future. I think there are far less gotchas now regarding weird architecture than there were in the 2010s.