CES 2025 Thread (AMD, Intel, Nvidia, and others!)
- Thread starter iroboto
- Start date
Microsoft is adding "Cooperative Vector" support to DirectX which allows shaders to use tensor cores in the GPU, which sounds like it should work with all GPUs with tensor cores, but in the article it has wordings like "in NVIDIA’s new RTX 50-series hardware" so I don't know.
TopSpoiler
Regular
I expect all RTX GPUs will be able to use DXCV for the Neural Radiance Cache. Previously it used tiny-cuda-nn which relies on the CUDA runtime.
I mean, it's going be obviously with AI enabled. There's frankly just no way its' going to have the same amount of raw power.
I think people need to get used to this concept. We have no other direction to head at this point in time, raw power is cost prohibitive in the consumer market, console market, etc. And also, plainly, it's getting more and more difficult to scale up fidelity and resolution due to a huge number of bottlenecks. While these are AI generated pixels, they get the job done.
I don't think it's a misinformation from that perspective, we need a new medium in which to compare to, I don't think raw power is worth discussing any further.
DegustatoR
Legend
Exactly. Such approach has already lead us to some questionable comparisons like DLSS2+ vs FSR2+ at a similar scaling level - which has never produced comparable results in IQ to begin with. Framegen is here to stay and all the other AI based tech will inevitably gain more and more traction. Just ignoring that because other cards can't do it is a completely wrong way to compare the products.
Well we know that, but it's likely going to fool plenty of other people who dont know better.
Cuz it's not like Nvidia didn't use to actually do this. 970 = OG Titan. 1070 = Maxwell Titan.
Now they're selling us a 5060 renamed as a 5070 and telling us it'll perform like a 4090(aka Lovelace Titan) even though it'll probably be more like the 4070Ti in reality. I would definitely rank this well into the 'straight up dishonest' category rather than just some cheeky marketing.
We would need to wait for benchmarks before judgement.
I mean, we can wait for benchmarks to know for sure, but we can definitely make some pretty clear educated inferences here.
Not even ray tracing? Just straight up, the ONLY thing worth talking about anymore is AI?
Maybe we can talk like this 10 years from now or something, but general rendering power still matters massively today.
Well, AI is being used to improve Ray Tracing, so having more AI power would improve RT performance in this case.
If the 5000 series is a bridge too far for folks, the 6000 series will be significantly more AI based. Nvidia currently is playing a leadership role defining where computing is headed, and unless we have a breakthrough in silicon advancement, AI is the only piece of software kit that can provide super high computational ability at a fraction of the power and silicon cost.
It's hard to imagine another IHV going in another direction and obtaining better performance than AI generated pixels at the same power, performance, price points.
It still matters a lot but it matters less for the most advanced elements of rendering. All of the hard stuff (GI, shadows, reflections) need some sort of ray querying engine unless you want to go back to baked lighting which is an option if you don’t mind static environments.
I would love to see a new game go hard using classic rendering techniques so we can have an actual counterpoint but I doubt that’s going to happen because those techniques are fundamentally limited in what they can achieve. It’s not worth it.
Well I'm personally not understanding how 'ray tracing' capabilities are somehow not included within the 'raw power' aspect.
AI is improving ray tracing much like AI is improving rasterized graphics, so I dont think that's a good argument. Even Nvidia seems to think raw power for ray tracing still matters or else they wouldn't have bothered improving their ray tracing cores for Blackwell.
neckthrough
Regular
It's probably best to talk about the 3 categories separately:
1) Raw shader flops
2) Fixed-function acceleration of parts of rendering algorithm (RT etc.)
3) AI-based image generation (which needs specific fixed-function blocks)
As of today and the foreseeable future, the cost of a transistor isn't going to improve much. So the question is, what's the best way to split the precious transistor budget between the above 3 categories? For sure the biggest bang for the buck seems to be with #3 -- with caveats of course.
Imagine a hypothetical scenario -- what if the transistor budget for the 6090 is exactly the same as the 5090. No clock speed increases either. How would you re-distribute the resources on the GPU? Some may want to go all-in on #1. Others may want to reduce #1 and #2 in favor of more #3.
Raw power matters a ton for RT. We’re at the very early stages of ramping up raw RT throughput.
Yeah. We need a lot more rays to feed the magic AI.
I dont think cost of transistors not improving is the entire reason that we cant have better value GPU's(or rather, better improvements in performance per dollar over time than we're currently getting), but that's a headache of an argument to deal with on this forum where 'greed' is basically not a concept that exists or could even theoretically exist.
But ignoring all that, my real point was that 'raw power' still matters, be it rasterization or ray tracing power. Even if its role slowly declines over time, it's not anywhere close to being completely irrelevant yet, which is what the original claim was saying. I was merely disagreeing with the general extremeness of the argument.
The problem is that while shader performance is easier to quantify (although there were some caveat but fortunately today most shader unit designs are quite sane), it's much harder to quantify raw "ray tracing" performance. I mean, what does the "318 TFLOPS ray tracing performance" means? How is that calculated? Can you compare it directly to other vendors' numbers? Probably not. This makes such numbers very marketing oriented. From this regard, the so called "AI TOPS" is actually better because it's actually easier to quantify (although because there's current no standard between vendors so it's still difficult to compare them, e.g. what does the "TOPS" mean? FP4? INT8? Who knows?).
MOD MODE: I love the conversation about technically evaluating AI upscaling performance, so I slung it out into this thread here: https://forum.beyond3d.com/threads/spinoff-technical-evaluation-of-upscaler-performance.63915/ It was hard to decide where to split it, so I left all the conversation about "what is the better use of transistors" in here.
ALSO MOD MODE: This is to nobody in particular, but also everybody: please make sure to keep the conversation rooted in your perspectives on the technology rather than your differences with the people in the thread. It's getting a little warm in here, and honestly I would expect some vigorous discussion on this topic of where to spend transistors because there will never be one right and true answer. Just make sure to stick to the topic and keep way from being angry at anyone; they have a right to their opinion just as you have a right to yours.
Back to just me posting: I suspect the raw RT power budget will continue to increase, only because I doubt anyone thinks "we have this solved now." At the same time, I'm under no illusions that the move to more AI will continue in earnest for the foreseeable future. The whole reason DLSS and FG came about was as a nod to a general lack of undelying brute force to get "full frames" onto our screens quickly, and the transistor cost was cheaper to upscale and frame-generate rather than actually try to brute force it.
As for this whole "fake frames" thing -- I still don't agree with this stance any more than I disagree with DLSS. Let's start with DLSS tho: everyone should remember how early DLSS versions were plagued with artifacts, severe object ghosting under motion being the prime artifact to my memory, but others included unstable / swimming effect on otherwise regular patterns, or ringing artifacts along very thin high contrast edges or point lights. A few of those still exist depending on settings, however I'm of the opinion that DLSS 3 and later have largely resolved these issues. I'm very happy with modern DLSS, because it allows even my 4090 to deliver more frames per second at the very absurdly highest graphical settings, at measurably lower power, with (what I perceive as) excellent image quality.
Now let's bring it back to frame generation: with DLSS enabled even at a quality preset, you're still "generating fake frames", just in a fractional format. IIRC DLSS quality is 75% native (I bet I'm off by a tad, but whatever) so you're already generating 33% "fake frames" (remember the math would be the percentage UP from native, so 75% to 100% is a 33.333% bump...) Using the DLSS balanced mode I think is 50% of native, so literally HALF of all displayed frames are "fake" at this point. Those "fake frames" are still subject to the same possibilities for creating artifacts for items coming on screen (eg the "real" raster is missing the new object by a pixel or two, but the upscaled image generates those pixels without knowing there should be an object present...) Funny thing is, NVIDIA's framegen is still using the same underlying detail as DLSS in terms of motion vectors, so it still understands how things are moving in a depth-parallax sense. Is it prone to more artifacts, like DLSS v1 was? You bet. Is it going to continue to get tuned over the next five years, like DLSS 1 was? Yup, it certainly is.
So, if you think you hate "fake frames", please at least consider remaining logically consistent and disable DLSS. Thanks
ALSO MOD MODE: This is to nobody in particular, but also everybody: please make sure to keep the conversation rooted in your perspectives on the technology rather than your differences with the people in the thread. It's getting a little warm in here, and honestly I would expect some vigorous discussion on this topic of where to spend transistors because there will never be one right and true answer. Just make sure to stick to the topic and keep way from being angry at anyone; they have a right to their opinion just as you have a right to yours.
Back to just me posting: I suspect the raw RT power budget will continue to increase, only because I doubt anyone thinks "we have this solved now." At the same time, I'm under no illusions that the move to more AI will continue in earnest for the foreseeable future. The whole reason DLSS and FG came about was as a nod to a general lack of undelying brute force to get "full frames" onto our screens quickly, and the transistor cost was cheaper to upscale and frame-generate rather than actually try to brute force it.
As for this whole "fake frames" thing -- I still don't agree with this stance any more than I disagree with DLSS. Let's start with DLSS tho: everyone should remember how early DLSS versions were plagued with artifacts, severe object ghosting under motion being the prime artifact to my memory, but others included unstable / swimming effect on otherwise regular patterns, or ringing artifacts along very thin high contrast edges or point lights. A few of those still exist depending on settings, however I'm of the opinion that DLSS 3 and later have largely resolved these issues. I'm very happy with modern DLSS, because it allows even my 4090 to deliver more frames per second at the very absurdly highest graphical settings, at measurably lower power, with (what I perceive as) excellent image quality.
Now let's bring it back to frame generation: with DLSS enabled even at a quality preset, you're still "generating fake frames", just in a fractional format. IIRC DLSS quality is 75% native (I bet I'm off by a tad, but whatever) so you're already generating 33% "fake frames" (remember the math would be the percentage UP from native, so 75% to 100% is a 33.333% bump...) Using the DLSS balanced mode I think is 50% of native, so literally HALF of all displayed frames are "fake" at this point. Those "fake frames" are still subject to the same possibilities for creating artifacts for items coming on screen (eg the "real" raster is missing the new object by a pixel or two, but the upscaled image generates those pixels without knowing there should be an object present...) Funny thing is, NVIDIA's framegen is still using the same underlying detail as DLSS in terms of motion vectors, so it still understands how things are moving in a depth-parallax sense. Is it prone to more artifacts, like DLSS v1 was? You bet. Is it going to continue to get tuned over the next five years, like DLSS 1 was? Yup, it certainly is.
So, if you think you hate "fake frames", please at least consider remaining logically consistent and disable DLSS. Thanks