The G in GPU is for Graphics damnit
ut21.github.io
ut21.github.io
Basically, almost half of their revenue is pure profit and all of that comes from AI.
While the slide looked a lot nicer, the data is also available on their site https://nvidianews.nvidia.com/news/nvidia-announces-financia...
When more of their revenue comes from AI than graphics, and they're literally removing graphics output from their hardware...
NVIDIA basically owns the market because of the stability of the CUDA ecosystem. So, I think it might be fair to call them an AI company, though I definitely wouldn't call them just a hardware maker.
You aren't even "dying on this hill", people like you are inventing a hill made out of dead bodies.
A company selling knives is not considered a butcher or cook, despite the main uses of knives being just that.
Nvidia spends a lot of money investing in downstream AI companies, in what feels like a rather incestuous circle
Next up: quantum. And that will be the end of them.
And other than maybe the crypto stuff, luck had nothing to do with it. Nvidia was ready to support these other use cases because in a very real way they made them happen. Nvidia hardware is not particularly better for these workloads than competitors. The reason they are the $4.6T company is that all the foundational software was built on them. And the reason for that is that JHH invested heavily in supporting the development of that software, before anyone else realized there was a market there worth investing in. He made the call to make all future GPUs support CUDA in 2006, before there were heavy users.
At some point, maybe it isn’t luck anymore but a general trend towards parallel computing.
> parallel computing.
"No, I see the pee" and at least another that I'd rather not express in polite company ))
They quietly (at first) developed general purpose accelerators for a specific type of parallel compute. It turns out there are more and more applications being discovered for those.
It looks a lot like visionary long term planning to me.
I find myself reaching for Jax more and more where you would have done numpy in the past. The performance difference is insane once you learn how to leverage this style of parallelization.
https://docs.jax.dev/en/latest/jax.numpy.html
A lot of this really is a drop in replacement for numpy that runs insanely fast on the GPU.
That said you do need to adapt to its constraints somewhat. Some things you can't do in the jitted functions, and some things need to be done differently.
For example, finding the most common value along some dimension in a matrix on the GPU is often best done by sorting along that dimension and taking a cumulative sum, which sort of blew my mind when I first learnt it.
In a way it's the scientific/AI/etc enterprise use of Nvidia hardware that enables the sale of consumer GPUs as a side effect (which are just byproducts of workstation cards having a certain yield - so flawed chips can be used in consumer cards).
Source (I am not sure how reliable this is because I got this from ChatGPT, but I remember seeing something similar from other sources): https://www.fool.com/investing/2024/02/12/gaming-was-nvidias....
Not as technicial but similarly cool.
funny side note. SEGA invested $5m in Nvidia then, after the fiasco to keep them alive. They sold that equity when Nvidia went IPO for roughly $15m. Have they kept it, it would be worth $3b today. SEGA's market cap is around $4b today.
So what follow is an endless series of tricks. Path-tracing is one of the purest implementations, and it is actually a simple algorithm to implement, but if you don't want to have a noisy mess on all but the most simple shapes, now we are talking PhDs and rock star developers.
Right now, I'm heavily into Cyberpunk 2077. I've got an RTX 5090, so I can turn all the details, including the ray tracing, to their maximum settings. It's absolutely gorgeous (Especially on my 4K HDR OLED monitor), but if you look really closely, you can still see the evidence of some shortcuts being taken.
Some reflections that are supposed to be a bit rough (like a thin puddle in the road) may appear a bit blurry as I'm walking, but will come into better focus when I stop. My guess is that as I'm moving, the angles of the rays being reflected change with every frame, making the data very noisy. Once I stop, they become consistent, so the reflection becomes clear.
Back in 2014, Disneys Hyperion farm supported 10-20 software bounces! [1] Does this count, or does it still require "cheats"?
[1] https://www.engadget.com/2014-10-18-disney-big-hero-6.html
And reflections and refractions.
Raster-based reflections are a simple shortcut: Just take the rendered image of what's being reflected and invert it.
But that doesn't work when the reflected object is off screen. As a result, if you're over water that's reflecting a city skyline or something in the distance, then pitch the camera down, the reflection vanishes as the skyline goes off screen.
Alternatively, you can create an environment-mapped texture, but that makes the reflection not reflect what's actually there, just an approximation of it.
I find it incredibly distracting in games. It's like bad kerning: Once you know what it looks like, you see it EVERYWHERE.
And I'm not sure that space has been economical for a long time. Integrated GPUs have more-or-less reached a point where they can handle PC games (albeit not at the latest-and-greatest resolutions/frame-rates/ray-tracing/etc), and the market for multi-thousand-dollar dedicated GPUs just isn't very big
What market research underpins this?
That's a rough upper bound of 20 million on the whole market, and NVidia + AMD already have it buttoned up - a newcomer can expect to attract a tiny fraction thereof
But I think more importantly, you can see this in NVidia's focus. Most of their profits are not in the gaming space anymore.
Look up the MSRPs of the most common GPUs. Very very few are over $1,000. If you want to interpret "multithousand" to be >= $2,000, then the answer becomes VERY small (less than 0.5%), as the highest MSRP gaming GPUs are the RTX 5090 and RTX 3090 Ti, both which technically have an MSRP of $1,999, that typically only applies to the "Founders Edition" releases done by NVIDIA. 3rd party AIBs (MSI, Gigabyte, Zotac, etc) typically charge a bit more.
[0] https://store.steampowered.com/hwsurvey/Steam-Hardware-Softw...
There is plenty of competition there: Qualcomm, Samsung, Apple, MediaTek and of course Intel and AMD, and things are moving fast. The best phone APUs nowadays are more powerful than my not so old MacBook Air M1.
> The best phone APUs nowadays are more powerful than my not so old MacBook Air M1.
Which is, itself, an APU.
The question is, is it better than a 2020 era dGPU and CPU combo (at any thermal/power envelope).
The answer is complicated unfortunately, but a 3090 (a 5 year old card) has 4x the memory bandwidth of an M4 Pro and also about 4x the FP32 performance.
So on the high end, descrete graphics cards are still going to be king for gaming. (I know that a 3090 isn't common, but 5080s are more powerful than 3090s).
PC gaming is a niche which is incredibly small. Ordinary people don’t use games on their PC provided they have one in the first place. Most PCs nowadays are laptops and they are mostly bought by companies sometimes by people and mostly to do work.
If you look at the respective market size, gaming is mostly done on smartphones and dedicated consoles and they all use APUs.
https://www.gamesindustry.biz/ign-launches-gaming-trends-pla...
> The prominence of mobile among younger players probably won't be a huge surprise to anyone reading this – 93% of Gen Alpha prefer playing on mobile, according to IGN's segmentation study. But preference for mobile is actually growing for Millennials, too, with 32% calling it their preferred device.
> ...
> Daily concurrent user numbers have grown in Roblox from 3.8 million in June 2022 to more than 25 million in June 2025. Over the same period, Fortnite has grown from 1.2 million to 1.77 million concurrents – with occasional blips, like when 15.3 million players logged on for the Marvel Galactus event.
Steam charts: https://store.steampowered.com/charts show 36,647,144 online now (as I write this)
I kinda feel that most games on smartphones are so fundamentally different to like the sweaty PC-gamer type games that they really should be considered a different market.
Take a look at the statistics for Minecraft and Fortnite, both games I would consider typical PC games, both massively successful. Mobile is always between 45% and 50%. PC has between 25% and 30% roughly on par with console.
PC gaming is mostly an Asian thing nowadays entirely propped up by esports. The market sure is big enough for GPU still making sense as a product (my incredibly small comment is admittedly a bit too extreme) but probably not for someone to go try to dislodge the current duopoly unless they have a product "for free" as an offshoot of something else.
I'm therefore curious. What do you think PC gaming is?
Could you render the scene in the simplest most efficient way possible, then train a special model that takes that picture, along with the coordinates of lights, depth maps, text descriptions of materials, etc, and adds detail?
DLSS - Render at low resolution, use an ML model to upscale
Frame Gen - Render a frame, use an ML model to generate the next frame. Render the next, gen the next....so on
(I think in practice, Frame Gen is more complicated due to fluctuation between FPS but that's another can of worms)
Anyway, I think you should change friends. Or study subjects. Either way.
Btw, these look like parasites or worms under microscope.
These friends don't get it!
But...with today's world of pixel shaders (Really, a world that's existed for 10+ years now), I'd be surprised if there's actually any benefit to be had these days. With a proper pixel shader, I doubt there's anything you could do on a CPU that you couldn't do on a GPU, and the GPU would be massively parallel and do it much faster.
But to play devil's advocate against myself, I have heard that programming for GPUs can be harder for many things. So maybe usability and developer-friendliness is what is meant by CPUs being better?
Basically, GPUs execute instructions in lockstep groups of threads. Each group executes the same instruction at the same time. If there's a conditional, and only some of the threads in a group have a state that satisfies the condition, then the group is split and the paths are executed in serial rather than parallel. The threads following the "true" path execute while the threads that need to take the "false" path sit idle. Once the "true" threads complete, they sit idle while the "false" threads execute. Only once both threads complete do they reconverge and continue.
They're designed this way because it greatly simplifies the hardware. You don't need huge branch predictors or out-of-order execution engines, and it allows you to create a processor with thousands of cores (The RTX 5090 has over 24,000 CUDA cores!) without needing thousands of instruction decoders, which would be necessary to allow each core to do its own thing.
There ARE ways to work around this. For example, it can sometimes be faster to compute BOTH sides of a branch, but then merely apply the "if" on which result to select. Then, each thread would merely need to apply an assignment, so the stalls only last for an instruction or two.
Of course, it's worth noting that this non-optimal behavior is only an issue with divergent branches. If every thread decides the "if" is true, there's no performance penalty.
Even that is more about control over the rendering process than what silicon is doing the work. With a lot of (older?) graphics APIs, you're implicitly giving the GPU's driver a lot of control over the output and how it is generated. This is how we got events like [0] where GPU vendors would make their drivers 'cheat' on benchmarks by trading image quality for speed when certain criteria were detected.
I imagine that tradeoff has changed somewhat as the industry has moved towards graphics APIs intended to give the programmer more direct control of the hardware.
[0] https://www.reddit.com/r/quake/comments/168btin/does_anyone_...
The simple calculations typically used for rendering graphics can easily be parallized on the GPU, hence it's faster. But the result should be identical if the same calculations are done on the CPU.
Also GUI frameworks like iced and egui typically support multiple rendering backands. I know iced is renderer agnostic, and can use a number of backands including the GPU graphics APIs Vulkan, DX12 and Metal.