If you aren't saying that, then why waste time engaging in sophistry, defending any part, of his pet conspiracy theory.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Rumor : AMD Bringing Ray Tracing Support to Navi in December
- Thread starter Sabrewulf...
- Start date
If you aren't saying that, then why waste time engaging in sophistry, defending any part, of his pet conspiracy theory.
If that’s your benchmark for proof then I can make the same claim about the rasterizers, texture units or any part of the GPU that has nothing to do with raytracing. Because tensors have nothing to do with it either.
Do you have any evidence of this? The Titan V does 110 TFlops vs Titan RTX’s 130 TFlops of peak 16 bit tensor performance. Yet the Titan V is clearly still faster than Titan RTX in some tensor workloads and equal in others.
It’s fascinating that you’re giving credibility to baseless uninformed speculation while questioning actual facts.
OutOfPhase
Supreme [H]ardness
- Joined
- May 11, 2005
- Messages
- 6,579
I'd suggest we go with Occam's Razor here. NV indicated they designed hardware specifically to accelerate RT. This hardware does actually have higher RT performance than everything else by a considerable margin.
The simplest possible explanation - they really did build RT HW.
The simplest possible explanation - they really did build RT HW.
BrotherMichigan
Limp Gawd
- Joined
- Apr 8, 2016
- Messages
- 376
You could take this even further and question why they would lie. I don't see that they have anything to gain by claiming they have dedicated hardware for ray tracing as opposed to saying, "We're leveraging other existing hardware to offload this task from the shaders."
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
Clearly they are jebaiting AMD into producing their own RT hardware...
The tensor cores in Turing are drastically different from the ones in Volta. Volta is also what NV used to demo and do early RT work on, cause yes it had the right hardware just missing the all important lower precision modes. Their star wars demo ect where running on Volta cards.
Volta can do Tensors at FP32 and FP16.
Turing adds FP8 and FP4.
Calculating rays doesn't need FP64 or FP32 or even FP16. 8 and even 4 bits of data is enough space to calculate one ray.
So yes the 2060 should be much better vs even a high end Volta card. It may have 1/2 the tensor cores but it can operate at 1/4 precision which is what Ray generation needs.
Now we can argue back and forth that their is a special little bit of silicon that is calculating those rays or that it is simply a Turing tensor core running at FP8 or FP4.... no doubt that either way that data is then feed into tensor cores for "denoising". We know that cause NV says so for sure. We don't have to argue about that. lol
My point the whole time is we are arguing semantics really. It doesn't really matter if NV has added a couple of hard silicon ray calculators... or if they are simply engaging clusters of their tensor cores to run at FP4 to calculate a metric F ton of rays. (its nothing but a marketing difference no one sees behind the driver anyway) That data is all feed through the tensor unit regardless of how its generated... cause DXR still isn't calculating enough rays to provide a noise free image.
Does it really matter... if a CCX of their tensor hardware is operating at FP4 and calculating rays, or if its a fixed function Matrix math processor which seems sort of dumb as that is what tensors are designed to do. It gets feed through the tensor cores to do denoising per NV marketing anyway. lol
And again for the record I am not anti NV. They are first to market with a consumer GPU that can do passable real time ray tracing. Ya other companies have dabbled and built a few demos... but NV has delivered. It may not have been a smooth launch and even today there are not a lot of games using DXR. Still the tech works and especially if you have an unlimited GPU budget there is not doubt there are a few examples of it looking fantastic in shipping games. AMD got beat to the punch no doubt, we all agree. I simply don't believe that NVs solution of using a combination of "RT cores" or "FP4 tensors" and FP16 tensors to denoise output is the winning solution long term. Personally I believe AMD is going to use a shader setup... but I could be wrong we'll know soon enough when they launch big navi it will have or won't have dedicated RT hardware. Intel is also about to enter the market and their solution may be either both or neither.
It will be nice to have competition in the RT field. I do really hope AMD and Intel have decent solutions that push the tech forward. It is promising. In a few years I think the RT stuff is going to look very different. I wouldn't even be surprised if after NVs first 7nm parts they move to a chiplet type part and remove tensors from consumer parts completely. There is a reason RTX cards got more expensive... those chips full of non GPU bits are not cheap to fab.
For the same reason some people get upset here when you suggest they didn't build Turing only to power gaming cards.
NV marketing believes, rightly so... that they need to say designed for gamers. Cause its the way its meant to be played after all.
People where already annoyed they had to sit so long on their 1080s.... Volta came and people got excited and NV did not use them for consumer grade cards. They skipped an entire cycle as they could. (we can blame AMD for that) Why put really expensive massive Volta chips into consumer parts if your still selling tons of 1070s and 1080s.
So with Turing they made it sound like the wait was long as they where working on the ultimate gamer card... REAL TIME RAY TRACING for realzies. It was a project years in the making... and yet somehow no game developers where really working that hard on it. Considering how much money NV has lavished on developers to get them to roll out things like Hair works I find it interesting that this years in the making clearly super expensive massive GPU got such crap software support from the kings of developer support programs. Yes that is a back handed compliment. NV does have perhaps the best developer support network in the industry and they had no games ready on launch day... and even today one year post launch their are SEVEN games on the market with RTX features, and one of them is quake 2.
They lied cause they needed to hype 2080s... they aren't really any faster then 1080s but they have this really cool new feature which is awsome. Now RT is going to be awsome. Reality is though most people that bought a 2080 launch day where likely expecting more then 1080ti performance in 99% of their games and 7 games capable of using their new killer feature a year after launch. 2080 is king sure... but really so far I have only seen one game (control) myself where I can honestly say I can even tell RT is on.
This is false. Turing adds INT4 and INT8, not FP8 and FP4.
Yes ray intersection arithmetic can be done using lower precision floating point units. However both Volta and Turing tensors only go down to FP16.
Turing has no tensor advantage here.
Indeed its late and I mistyped.
However we are now really really stretching.
https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/
"Turing Tensor Cores
Tensor Cores are specialized execution units designed specifically for performing the tensor / matrix operations that are the core compute function used in Deep Learning. Similar to Volta Tensor Cores, the Turing Tensor Cores provide tremendous speed-ups for matrix computations at the heart of deep learning neural network training and inferencing operations. Turing GPUs include a new version of the Tensor Core design that has been enhanced for inferencing. Turing Tensor Cores add new INT8 and INT4 precision modes for inferencing workloads that can tolerate quantization and don’t require FP16 precision. "
Ray tracing does not require a Float Point 16 level of precision, and a gamer doesn't care if a handful of the rays being calculated aren't 100% accurate. At the very least they are using an inference engine to do RT denoise, as they have said the tensors are used for denoise. That is no doubt using an inference engine... the scene has a set of data the engine uses a set of rules to denoise. No real debate here... that is what NV has said in marketing and white papers. From their own marketing yes... Turing has been enhanced for faster inferencing operation... the main enhancement, lower precision modes.
And none of that has anything to do with raytracing. You're certainly not doing intersection tests using int8.
Which raytraced games are using tensors for denoising? If I recall correctly they're all using their own denoising algos running on regular shaders.
And again, that's about as convincing as the Grand Wizard of the KKK, saying he isn't anti African American.
Protestation of innocence, matter very little,when you are engaged in an active disinformation campaign that claims NVidia is lying to everyone.
While your beliefs might be of interest to a psychologist, all you have done vociferously defending your nonsense claims, is show that you don't understand how technology, or business work, and that your opinions run counter to reality. So you have only succeeded in completely devaluing your opinions.
I also posted this AMD Ray Tracing slide up the page. AMD is crystal clear. Their RT Shader (RDNA gen 1) solution is NOT for games. They will have an RT Hardware solution for Ray Tracing on RDNA generation 2. That is direct, clear, and straight from AMD.
But then again you might be inclined to believe the voices in your head instead of AMD, just like you do for NVidia.
Don’t count on your 5700xt doing RT well if at all. Rdna2 cards coming 2020 that will.
RT by AMD is been announced for some time. Vega is supposed to be able to handle Raytrace in a usable way by using FP16. It means Ray Trace support will be better than on Pascal, but not on RTX cards on par with Vega on non RTX use, if AMD make it into the drivers.
So this is coming to Navi, but not sure AMD won't put it in Vega drivers, because it is just as much possible. Polaris series will be unable to handle RT in the way AMD was announcing it back in the days.
Now this is not real hardware support by specialized circuitry, but it is unclear if next RDNA 2 and Intel Xe will ever have specialized RT circuits. Nvidia Volta is able to handle those without specialized units but with more Cuda units.
RTX cards seem to be some kind of a rush by Nvidia to put big RTX support but it seems it is not a very efficient implementation and clearly not future proof. Nvidia needs to implement RTX functions into CUDA units.
Don't bet on your RTX future support !
So this is coming to Navi, but not sure AMD won't put it in Vega drivers, because it is just as much possible. Polaris series will be unable to handle RT in the way AMD was announcing it back in the days.
Now this is not real hardware support by specialized circuitry, but it is unclear if next RDNA 2 and Intel Xe will ever have specialized RT circuits. Nvidia Volta is able to handle those without specialized units but with more Cuda units.
RTX cards seem to be some kind of a rush by Nvidia to put big RTX support but it seems it is not a very efficient implementation and clearly not future proof. Nvidia needs to implement RTX functions into CUDA units.
Don't bet on your RTX future support !
Pascal support for RTX is relying on low speed integer functions on Pascal that are faster even on Turing GTX 1660 series (compared to GTX 1060 practically equal card), because they are still faster tan FP functions.
On the contrary, Vega and Navi may rely on FP16 fast units to speed up RT. Not on par with RTX but way faster than Pascal cards.
On the contrary, Vega and Navi may rely on FP16 fast units to speed up RT. Not on par with RTX but way faster than Pascal cards.
It's not 64 or 32 bits its 64 or 32 threads. Also, RDNA can perform 1 wave64 over 2 cycles or 1 wave32 every cycle, not what you mentioned. It uses an entire WGP to do so, which is 2 CU's -- so in terms of threads per cu per clock it's identical to GCN.
GCN = one wave64 over 4 clocks on a single 16-wide CU
RDNA = one wave64 over 2 clocks on a single WGU (two cu's, so 32-wide) -- or a single wave32 in a single clock on a single wgu.
funkydmunky
2[H]4U
- Joined
- Aug 28, 2008
- Messages
- 3,871
I don't think it matters. Whatever AMD and its console partners decide will become the de-facto standard.
Like many have stated, there is only a real half-dozen games supporting Nvidia's launch into the future after over a year. Not a viable tech unless your getting a pay check to offset the cost.
This is too many shades of Matrox bump-mapping, and its last gasp triple-head. A complete architecture designed to separate its self from the competition. They should have focused on FPS like the 1080ti did.
Like many have stated, there is only a real half-dozen games supporting Nvidia's launch into the future after over a year. Not a viable tech unless your getting a pay check to offset the cost.
This is too many shades of Matrox bump-mapping, and its last gasp triple-head. A complete architecture designed to separate its self from the competition. They should have focused on FPS like the 1080ti did.
Dr. Righteous
2[H]4U
- Joined
- Aug 1, 2007
- Messages
- 3,163
DirectX ? WTH???
Why are they not working with Vulkan RT? Even Nvidia is doing this.
Why are they not working with Vulkan RT? Even Nvidia is doing this.
Gotta start somewhere and I think all the shipping games with RT are DirectX.
Wolfenstein: Youngblood is supposedly getting RT at some point but it hasn't happened yet.
And maybe a massive wasted opportunity for nvidia. Had nvidia done ray tracing add in cards that worked regardless of gpu used, it would have been awesome.
But they decided to do this 'added feature, added price' strategy of no real value.. if only to move the price brackets upwards.
Seems most likely AMD just baked it in the architecture and called it a day.. as opposed to some extra rt cores or some such.
In the end, Nvidia and its billions.. so how can one argue too much?
But they decided to do this 'added feature, added price' strategy of no real value.. if only to move the price brackets upwards.
Seems most likely AMD just baked it in the architecture and called it a day.. as opposed to some extra rt cores or some such.
In the end, Nvidia and its billions.. so how can one argue too much?
I am not sure, if anyone posted this before, so I apologise if this a duplicate post:
The latest enCore RT demo gives World of Tanks players a taste of ray tracing, specifically shadows, without requiring NVIDIA RTX graphics cards. In fact, this implementation was made through Intel's Embree library (part of OneAPI) and is compatible with all DirectX 11 graphics cards.
According to Wargaming, the performance of enCore RT will be as good or better as the ray tracing implementations we've seen so far in games that support NVIDIA RTX, while you can expect the quality to be very close.
However, it should be noted that the developers themselves have admitted ray traced shadows will only be cast from tanks as the performance would have otherwise suffered.
https://wccftech.com/world-of-tanks...ows-on-all-dx11-gpus-made-with-intels-embree/
The latest enCore RT demo gives World of Tanks players a taste of ray tracing, specifically shadows, without requiring NVIDIA RTX graphics cards. In fact, this implementation was made through Intel's Embree library (part of OneAPI) and is compatible with all DirectX 11 graphics cards.
According to Wargaming, the performance of enCore RT will be as good or better as the ray tracing implementations we've seen so far in games that support NVIDIA RTX, while you can expect the quality to be very close.
However, it should be noted that the developers themselves have admitted ray traced shadows will only be cast from tanks as the performance would have otherwise suffered.
https://wccftech.com/world-of-tanks...ows-on-all-dx11-gpus-made-with-intels-embree/
Going by the quote below from the game developers, it looks like compute heavy cards such as Vega 56, 64, Radeon VII might be good at ray tracing the Tanks
It’s not done mainly by the CPU. It just uses multiple CPU threads to effectively rebuild high
quality BVH structure for every frame. Ray tracing itself is done using compute shaders on the GPU. The high-quality BVH structure allows for better ray tracing performance on the GPU
Was about to post this as well.
It’s actually a really good use case for RT as shadow maps need to be really high-resolution to accomplish the same level of self shadowing detail on objects.
Wouldn’t be surprised if next gen consoles take a similar approach for BVH updates with the massive CPU power at their disposal.
Dr. Righteous
2[H]4U
- Joined
- Aug 1, 2007
- Messages
- 3,163
Yeah; you are right. Just seemed a bit odd to me since it was AMD pointed out that legacy API was stiffing innovation; thus their work on Mantle and that lead to VulKan.
Shouldn't say AMD was alone in this. John Cormack talked about this same issue years before.
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
Dr. Righteous From what I understand, RT in Vukan is not a standard. Nvidia made their own extension, and I guess AMD (or Intel) can make an extension as well, but it's not standardized yet.
On Windows, though, DXR is a standard for DirectX12, which makes more sense to support now since there are shipping games using it.
On Windows, though, DXR is a standard for DirectX12, which makes more sense to support now since there are shipping games using it.
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
World of Tanks is interesting. It is a limited use of RT, but proves that dedicated hardware is not needed. Honestly looks pretty good.
Legendary Gamer
[H]ard|Gawd
- Joined
- Jan 14, 2012
- Messages
- 1,590
It will be somewhat meaty... Did any of you see those videos comparing Nvidia's RTX performance to the 5700XT in Minecraft with Ray Tracing enabled? While the RTX sat around 60 FPS the 5700XT was around 40. Now that's a pretty tangible hit but the 5700XT was still very playable. Granted Minecraft Ray Tracing is akin to a primitive screw head banging rocks together. So the RTX card was approximately 34% faster (if I didn't just fuck up my math).
Legendary Gamer
[H]ard|Gawd
- Joined
- Jan 14, 2012
- Messages
- 1,590
I was wondering about this. I have my AMD 3600 sitting in the closet after moving out of my house and I have been contemplating moving off my 9600 Intel to the 3600 and wondering how all my stuff was gonna play together. God knows I had a hellish time with my 2080Ti initially (on the third card, still somehow). Anyone else having a helluva time with the 2080Ti and AMD? Just curious, I guess I really won't know until I give it a spin.
IdiotInCharge
NVIDIA SHILL
- Joined
- Jun 13, 2003
- Messages
- 14,675
That's really the point of dedicated RT hardware, though -- in order to use RT everywhere, you're going to need some hardware. It's not like WoT is a graphically demanding title in terms of what's actually being rendered either. Take what we might expect of next years AAA games and then add RT (which should be the expectation regardless...), and then try to use a CPU hack.
And then try to do 4k.
And then try to do that with ~8ms frametimes for a consistent 120FPS.
We're several layers deep into 'nope'
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
Well, multi-core CPUs are underutilized in many games (maybe most of them). If RT was done on the idle cores we would be getting something essentially for free on current hardware.
IdiotInCharge
NVIDIA SHILL
- Joined
- Jun 13, 2003
- Messages
- 14,675
This assumes multicore > 6, which while solidly in enthusiast territory, is nevertheless quite uncommon for consumers, especially also at sustained clockspeeds that are useful and with useful IPC levels (thus excluding current consoles and mobile devices with 6+ cores).
And then, there's the issue of complexity: the software solution implies the assumption that it would be in addition to hardware raytracing, which adds complexity and pushes the implementation solidly into niche territory.
Now, if an engine were used that were designed with that complexity in mind, then cool -- assuming that the use of FP resources on the CPU doesn't have other unintended effects such as decreased clockspeeds due to higher TDP from additional resource usage or stalling threads due to additional cache, memory, and memory bandwidth usage, etc.
can't even get sli so get real.
OutOfPhase
Supreme [H]ardness
- Joined
- May 11, 2005
- Messages
- 6,579
Unfortunately, RT is quite heavyweight. I can't see how you'd get any timely output from a theoretically-idle CPU core. It's quite taxing/crippling on even idle conventional GPU resources which are far more suited to the task.
N4CR
Supreme [H]ardness
- Joined
- Oct 17, 2011
- Messages
- 4,947
Nothing does 4k120... In a modern title. Even with the hardware.That's really the point of dedicated RT hardware, though -- in order to use RT everywhere, you're going to need some hardware. It's not like WoT is a graphically demanding title in terms of what's actually being rendered either. Take what we might expect of next years AAA games and then add RT (which should be the expectation regardless...), and then try to use a CPU hack.
And then try to do 4k.
And then try to do that with ~8ms frametimes for a consistent 120FPS.
We're several layers deep into 'nope'
So that's a big nope all round till maybe 2022 or more..
Even then it is still only partial scene kludging of today.
Wot is just more partial scene than the rtx stuff already is.
I agree with all points. But like ChadD pointed out, RDNA is new and unique and I am excited to see what comes of this.
IdiotInCharge
NVIDIA SHILL
- Joined
- Jun 13, 2003
- Messages
- 14,675
Exactly -- WoT is a pretty shallow though cool implementation using an extremely limited technique, while titles that push GPU RT hardware to its limits are still not taking ray calculation to the point of pathtracing, and the performance needs to do that exceed what current GPUs can satisfy to hit 4k120.
Ergo, we shouldn't expect the addition of a handful of CPU cores tracing rays to make much of a difference
It proves dedicated hardware is not needed if you're doing very lightweight RT. We already knew that.
The World of Tanks implementation does BVH construction on the CPU and DX11 shaders for ray intersection. Very, very cool.
Looks like Turing is ahead by a significant margin without the help of its RT hardware.
https://wccftech.com/world-of-tanks...d-polaris-vega-navi-pascal-and-turing-tested/
crazycrave
[H]ard|Gawd
- Joined
- Mar 31, 2016
- Messages
- 1,884
I ran the benchmark in RT Ultra 1080p and recorded with phone as Relive holds the clocks back .