AMD raytracing patent

MangoSeed

[H]ard|Gawd
Joined
Oct 15, 2014
Messages
1,447
Interesting AMD patent recently published. No guarantee this is how RDNA will do raytracing but it gives a good idea of how AMD is thinking about it.

https://www.techspot.com/news/80724-amd-patent-describes-hybrid-approach-real-time-ray.html

34gqag5.jpg
 
Also a handy way to limit your competitions options with patents, Just like Nv with Tensors and such.
 
Not sure how useful this patent would be for limiting competition. It’s basically reading BVH data from a cache, passing it to some fixed function hardware and sending the result to the shader core. Pretty much what nvidia is doing.

The interesting part is doing scheduling on the shader comes and that’s software and software approaches aren’t patentable. The actual code is copyright protected but nothing is stopping someone from writing a software Ray scheduler on nvidia hardware.

In the end only performance matters.
 
This is very similar to what NVIDIA is doing on Pascal or None-RT core Turings...so I predict the performance will be very similar to DXR running on Pascal, aka slower than RTX cards by a factor...
 
This is very similar to what NVIDIA is doing on Pascal or None-RT core Turings...so I predict the performance will be very similar to DXR running on Pascal, aka slower than RTX cards by a factor...

It's talking specifically about using fixed function BVH hardware in some cases (and not using it when it doesnt need to) so it's nothing at all like Pascal because Pascal doesn't have fixed function BVH hardware. More like Turing most likely as it would be pretty naive to think that nvidia isn't using optimizations of some sort as well.
 
It's talking specifically about using fixed function BVH hardware in some cases (and not using it when it doesnt need to) so it's nothing at all like Pascal because Pascal doesn't have fixed function BVH hardware. More like Turing most likely as it would be pretty naive to think that nvidia isn't using optimizations of some sort as well.

The "fixed function" are TMU's...
 
The "fixed function" are TMU's...

Not quite. There is also a new hardware block for calculating ray intersections similar to Turing. That calc is the most expensive step in raytracing. The difference is that nvidia likely traces the ray through the entire BVH in one shot while the AMD approach does just one node before returning control to the shader core to setup the next pass. At face value that would mean AMD’s approach is slower but more flexible.

Hard to say if it’s actually that different until we learn more about nvidia implementation.
 
Not quite. There is also a new hardware block for calculating ray intersections similar to Turing. That calc is the most expensive step in raytracing. The difference is that nvidia likely traces the ray through the entire BVH in one shot while the AMD approach does just one node before returning control to the shader core to setup the next pass. At face value that would mean AMD’s approach is slower but more flexible.

Hard to say if it’s actually that different until we learn more about nvidia implementation.

If AMD and Radeon past history is anything to go by, it does not matter so much if the Radeon doing ray trace is a "bit slower per computation or however that is best worded, there is also the likely hood AMD was/is trying to ensure all them "not so quick" lil ray tracers are fed as much as they need as "smooth" as possible. (better to have many many cores fed their data in large enough "blocks" so they can work to the best of their ability...which drives multi-core mutli-data calcs through the roof done right (or the very least keeps the latency between frames etc much lower (as the "master" is not waiting for the ray trace bit to finish off.

word wall ^.^

Nv and Intel for EVER were pretty much focused on get per core and clock up as high as possible no matter the cost, which was great for a while, AMD and Radeon did VLIW design, which lead to Polaris - Vega - Navi .. per core per clock are "competative" no matter what fool says what, the major difference being they can of course do things the others cannot, they (AMD) seem to be saying

"Per core is "good enough" we need to get more cores to work together as best as possible for the lowest cost possible, as this is what is going to drive the future, not crazy speed per core, not one off feature only used for one or 2 applications, that was good enough for the past, we need to go beyond this or status quo remains so therefore what choice does anyone have?, now you have true choices no matter the budget or workload(s)..
 
Not quite. There is also a new hardware block for calculating ray intersections similar to Turing. That calc is the most expensive step in raytracing. The difference is that nvidia likely traces the ray through the entire BVH in one shot while the AMD approach does just one node before returning control to the shader core to setup the next pass. At face value that would mean AMD’s approach is slower but more flexible.

Hard to say if it’s actually that different until we learn more about nvidia implementation.

It does sound a bit like that, but really after reading all the NVidia stuff, it looks like it also does "Intersection test - Shading, Intersection test - Shading, etc...).

It really seems that it has to be that way. At each intersection, you need to gather the shading information, and carry it to the next intersection.

But you also have to consider we are comparing NVidia information for public consumption to an AMD patent.

So it helps to have to have a passing familiarity with how patents are written. I used to work in Telecom, and our company had Patent Bounties. Small rewards for starting patent application (subject to internal approval), bigger awards if the patent was awarded, and I saw the patents on a couple of simple software techniques that coworkers had awarded. A simple technique in a few lines of code appeared a lot more convoluted and complex, after the back and forth withe patent attorneys was done.

The whole point of Patent applications is to make something simple and obvious seem like some kind of complex, and unique work of genius to snow the patent inspectors.

The differences are likely minutia, that will really be impossible to determine which solution is really better overall. All we can judge in the end is how each individual product implementation delivers, and that will be about a lot more than the minor differences in the efficiencies of each design, since the lessor solution can just compensate with a bit more transistor budget.

IMO it will be interesting to see if AMD bothers with Tensor cores in it's design, as so far they look rather superfluous in RTX.
 
Interesting. There may be two new Lines of Nvidia cards before AMD comes with this idea on its Arcturus line aka pure RDNA cards with hardware RT implemented. It means, we'll see Nvidia Super made at Samsung on 11lpp, already 10/15% better than first Turing Line, then we'll see Ampere based on 7nm EUV, then we'll see the upgrad from Ampere at the rate Nvidia use to come with their new line every year, and then only we'll see first try from AMD.
Not sure AMD can catch up when their 7nm GPU cannot come close to Nvidia GPU which manages at the same time to make Raytracing already real.

Maybe we should stop looking at AMD which seems good on CPU now, and watch what Intel is able to bring. My thinking is that AMD never took seriously GPU market appart from console gaming which doesn't have to compete but only on the price. GPu at AMD was only an argument for selling CPU. Though it is their APU that made their sales until recently against Intel. Now they don't even need to bother about their APU, and Navi shows that.
Navi is a very poor upgrade from Vega. In fact it trades the compute for gaming when Turing adds all the specs together. It's still on 12nm now 11nm, and has better gaming, better computing (now better than AMD Vega) and brings hardware Raytracing AND IA Tensor cores. Nvidia is now lightyears ahead of AMD.
 
Turing Super is the exact same chips made as before, on TSMC 12FFN. Where are you hearing Samsung 11nm ?

Navi is a very poor upgrade from Vega. In fact it trades the compute for gaming when Turing adds all the specs together. It's still on 12nm now 11nm, and has better gaming, better computing (now better than AMD Vega) and brings hardware Raytracing AND IA Tensor cores. Nvidia is now lightyears ahead of AMD.

Fanboy enough? Sheesh. What do you mean Navi is a very poor upgrade from Vega. Navi cards are now able to pretty much match the performance of nvidia cards with the same amount of shaders. That's a huge leap. WHat do you mean it trades the compute for gaming? Vega 10 wasnt really a compute chip anyways -- it has 1:16 FP64 perf. Vega 20, on the other hand, has more FP64 perf than anything except V100. In any case, Vega 10 is winning on all but one of the compute benches here even vs Turing Super. Sure Turing has Tensor cores but they are largely pointless for gaming. You arejust throwing around the term 'computing' and don't seem to really know what you're talking about. Light years ahead? Not really ...
 
Turing Super is the exact same chips made as before, on TSMC 12FFN. Where are you hearing Samsung 11nm ?



Fanboy enough? Sheesh. What do you mean Navi is a very poor upgrade from Vega. Navi cards are now able to pretty much match the performance of nvidia cards with the same amount of shaders. That's a huge leap. WHat do you mean it trades the compute for gaming? Vega 10 wasnt really a compute chip anyways -- it has 1:16 FP64 perf. Vega 20, on the other hand, has more FP64 perf than anything except V100. In any case, Vega 10 is winning on all but one of the compute benches here even vs Turing Super. Sure Turing has Tensor cores but they are largely pointless for gaming. You arejust throwing around the term 'computing' and don't seem to really know what you're talking about. Light years ahead? Not really ...

TPU took out the heat sink and saw the GPU is made from South Korea for RTX 2060 Super. Since there is no TSMC fab in South Korea, probably save to say Samsung produce it, though it is just my opinion. The RTX 2070 Super is still made from Taiwan.

Edit, Factum show me a post by Snowdog that is probably just packaging done in Korea.
 
Yeah Anandtech's review is still showing TSMC 12FFN for all the super chips. I would have imagined that there would be a fair bit more news about it if they changed the manufacturing process on the chips.
 
It also shows why rumor-sites are the bane of technology...so much fake shite spread by a few useless sites...
 
Can't wait to see what sammy silicon manages.
 
It also shows why rumor-sites are the bane of technology...so much fake shite spread by a few useless sites...

actually i'd say the half ass news sites that regurgitate information from known rumor sites that pull shit out of their ass are the bane of media. at least those people make an attempt, the copy paste news sites put no effort into what they post for a quick buck.
 
actually i'd say the half ass news sites that regurgitate information from known rumor sites that pull shit out of their ass are the bane of media. at least those people make an attempt, the copy paste news sites put no effort into what they post for a quick buck.

The outcome is the same...loads of false rumors, to wild expectations and garbage flooding forums...It's effing boring to watch, even if I never click links to those FUD sites, muppets keeps sharing them like they were the second comming and polluting forums with retard fake rumors.
 
The outcome is the same...loads of false rumors, to wild expectations and garbage flooding forums...It's effing boring to watch, even if I never click links to those FUD sites, muppets keeps sharing them like they were the second comming and polluting forums with retard fake rumors.

Here is a simple idea then, if you dont like the site or rumor then dont post on it and we dont have to listen to you whine about it, we all win. Rumors are great for seeing what others think of them good bad or indifferent, hell no site has ever been 100% on rumors anyway.
 
Sorry, but most websites show all Nvidia produced at TSMC which is officially false. Some forgot lower end Pascal were produced at Samsung. This had nothing to do with packaging. Nvidia just mentioned they were producing some chips at Samsung facility, never said to my knowledge, what they were producing until some people discovered their 1050 chip was made by Samsung.
This is not big deal. Samsung and Nvidia already do business together. No big news. Though, if it's 11nm Samsung it may be interesting since it is not exactly the same as 12nm TSMC. But however as a matter of fact, new Turing chips are better and will receive also faster GDDR6, which makes them around 10/15% better.
 
Here is a simple idea then, if you dont like the site or rumor then dont post on it and we dont have to listen to you whine about it, we all win. Rumors are great for seeing what others think of them good bad or indifferent, hell no site has ever been 100% on rumors anyway.

It's not the rumors themselves that are the main issue, it's the large number of people that don't examine them with a proper dose of skepticism, don't do any research themselves, and don't check the track record of the rumor site, before jumping straight to belief.

I like sifting through rumors and doing a bit of research to speculate on what is going to happen, but when you point out something reasonable, only to be met by a dozen people claiming that the Navi Part equal to RTX 2070 is going to be $250 (despite all reason indicating contrary) it gets tiresome, the same with $99 Ryzen 3000 6 cores, or whatever. Once there is a rumor some people refuse to believe anything BUT the rumor.

The same with this 11nm Samsung TU 106 rumor started by TPUP. It's nothing anyone told them. They just saw a Korea marking, and made up the rest.

Neither they nor their readers take the time to do 5 minutes of research in Google to notice that the entire RTX lineup has Korea markings on early demo/prototype cards from the beginning (2018), so it appears at some stage of prototyping, or early low production runs pass through Korea, and the chips get a Korea marking, but later full production cards all seem to be marked Taiwan.

It seems a LOT more likely that this is just packaging marking step, and not that NVidia prototypes with Samsung foundries, and then builds them with TSMC. Also not a sign that 2060 Super now uses Samsung die given the obvious Korea markings on past RTX cards.
 
It's not the rumors themselves that are the main issue, it's the large number of people that don't examine them with a proper dose of skepticism, don't do any research themselves, and don't check the track record of the rumor site, before jumping straight to belief.

I like sifting through rumors and doing a bit of research to speculate on what is going to happen, but when you point out something reasonable, only to be met by a dozen people claiming that the Navi Part equal to RTX 2070 is going to be $250 (despite all reason indicating contrary) it gets tiresome, the same with $99 Ryzen 3000 6 cores, or whatever. Once there is a rumor some people refuse to believe anything BUT the rumor.

The same with this 11nm Samsung TU 106 rumor started by TPUP. It's nothing anyone told them. They just saw a Korea marking, and made up the rest.

Neither they nor their readers take the time to do 5 minutes of research in Google to notice that the entire RTX lineup has Korea markings on early demo/prototype cards from the beginning (2018), so it appears at some stage of prototyping, or early low production runs pass through Korea, and the chips get a Korea marking, but later full production cards all seem to be marked Taiwan.

It seems a LOT more likely that this is just packaging marking step, and not that NVidia prototypes with Samsung foundries, and then builds them with TSMC. Also not a sign that 2060 Super now uses Samsung die given the obvious Korea markings on past RTX cards.
Forget about it. It is now known how Nvidia RTX Super got much better specs and that's beyond 15%. When they used a cut down GPU (2060), now they use these same nearly full GPU at much higher speed (2060S nearly the same as 2070). When they used full GPU (2070), now they use the GPU of the better card (2080) with a very mild cut down (2070S) at higher clock. When they were limited in memory (2060), they got full 8GB (2060S), when they already had the top GPU (2080), it is now completely fully used at higher clock (2080S) and with much faster Vram (16Gbps instead of 14). In fact you get something like the better card for the price of the lower spec card and keeping the old name+super.

In fact AMD Navi is left with no advantage at all.
All those Super cards will be sold at quite the same official price as the Navi 5700 and 5700XT, except that Navi cards were competing the old RTX cards at the same price and without the raytrace and VRS and other great RTX Features.
Furthermore Nvidia cought up with AMD on Turing regarding compute, but AMD used to sell their cards a bit cheaper. But Navi cards are much worse than Vega cards regarding compute, so AMD has even lost this battle by trying to make gaming only cards.

So there is no other interest than much much lower price that would make you buy Navi cards, but those have been announced at the same price as old RTX cards. So if AMD doesn't sell Navi 5700 XT at 250$, and 5700 at 180$ as it was supposed to in the leaks in january, there is NO interest in those cards and I mean really 0 interest. Even for mining you should get a Turing card... or a Vega card as long as AMD still sell them for a good price.
And mind that I WAS a great fan of AMD. I have a Vega 56 and a RX 580 Nitro, who are really great cards I would recommend eve against any Turing for the price, and I would even, sponsor Radeon VII against any Turing even Super, but stay away from Navi at the price they announced it. It's just double the price they should sell them.
Navi is a mistake ! And I bet AMD realized that and they try to gamble on some scam, like Nvidia did with Maxwell, but it won't work.
 
Scam by meaning that it's better than what meets the eye. They already told us that the Tflops are low but it doesn't matter (just don't look at those specs they say) because the card is optimized (?!).
Scam by meaning they are improving GCN for now and pure RDNA later wich will be great. The loss of compute power is not great. In fact the card is all made for gaming only and not any gaming but the games that will be optimized for that card. And the card is already not that good in DX12. Meaning : it doesn't have the brute capability to handle what is thrown at it without wrapping it through proprietary APIs. This is just how Nvidia Maxwell worked.
Means also that when AMD will stop optimizing his card for new games the card will look like older Nvidia cards, completely junk, while older AMD GCN cards are still competing today.
If I buy an AMD card it is not because I want the same thing as Nvidia does. Everything that was good in AMD cards is thrown away, to become some kind of a subequal to Nvidia.
Not interested. If I have to chose, I will chose the original. Nvidia.
 
You know the Tfops is calculated by just multiplying the number of shaders * 2 * clockspeed? It has NOTHING to do with the architecture?

The loss of compute power is because there are less shaders dude. NOTHING ELSE.

You are coming into this thread just trying to shit all over AMD but you clearly don't even know what you are talking about.
 

Nvidia’s hardware raytracing patent.

http://www.freepatentsonline.com/9582607.html

- Full tree traversed in hardware for each ray before returning intersection result to the shader core
- Multiple rays processed in parallel
- Dedicated memory (cache, stack, local storage) for ray traversal, separate from SM memory
- Simple compression scheme used to reduce size of tree structure in memory

Seems like a straightforward implementation of a binary tree traversal with optimizations for raytracing data structures. There’s opportunity to increase performance by adding more traversal units, larger cache etc in the future.

Given the amount of data movement and the complexity of the intersection algos it’s unlikely raytracing will be done on general shader cores anytime soon.
 
Given the amount of data movement and the complexity of the intersection algos it’s unlikely raytracing will be done on general shader cores anytime soon.
ekhm, and what bout Pascal 1060-1080Ti and Turing 1660 cards?
Q2-RTX-1080.png

It just works! ~_^

EDIT://
I bet Nvidia shader based algo is definitely not the best what can be achieved without dedicated hardware as it was more to be help for indie game developers to get to know technology and to make their 1660 series cards more attractive compared to competition which cannot do any RT with DXR at all. And also to make a point that dedicated RT cores are actually useful because some idiots thought it is just dead hardware, though this last point kinda makes spending too much effort on optimization less important compared to eg. just making it work and be as much compatible with RTX implementation as possible to avoid any software issues. This implementation is also showcase of Turing's improved CUDA engine.

With some clever tricks and additional (simple) hardware resources placed in right places and shader based ray-tracing might be made somewhat faster than that, especially if de-noising is moved out of Tensor cores to somewhere else with different less processing power demanding algorithm. I doubt however it will be significantly faster and definitely nowhere near full hardware RTX implementation
 
Last edited:
ekhm, and what bout Pascal 1060-1080Ti and Turing 1660 cards?
View attachment 171974
It just works! ~_^

EDIT://
I bet Nvidia shader based algo is definitely not the best what can be achieved without dedicated hardware as it was more to be help for indie game developers to get to know technology and to make their 1660 series cards more attractive compared to competition which cannot do any RT with DXR at all. And also to make a point that dedicated RT cores are actually useful because some idiots thought it is just dead hardware, though this last point kinda makes spending too much effort on optimization less important compared to eg. just making it work and be as much compatible with RTX implementation as possible to avoid any software issues. This implementation is also showcase of Turing's improved CUDA engine.

With some clever tricks and additional (simple) hardware resources placed in right places and shader based ray-tracing might be made somewhat faster than that, especially if de-noising is moved out of Tensor cores to somewhere else with different less processing power demanding algorithm. I doubt however it will be significantly faster and definitely nowhere near full hardware RTX implementation

NVIDIA also did it to show the benefit of RT-cores in the die....how much better RT-cores perform raytracing than shader cores ;)
Even some people to this day still claims there are no RT-cores in Turing 20x0 SKU's...
 
Interesting. There may be two new Lines of Nvidia cards before AMD comes with this idea on its Arcturus line aka pure RDNA cards with hardware RT implemented. It means, we'll see Nvidia Super made at Samsung on 11lpp, already 10/15% better than first Turing Line, then we'll see Ampere based on 7nm EUV, then we'll see the upgrad from Ampere at the rate Nvidia use to come with their new line every year, and then only we'll see first try from AMD.
Not sure AMD can catch up when their 7nm GPU cannot come close to Nvidia GPU which manages at the same time to make Raytracing already real.

Maybe we should stop looking at AMD which seems good on CPU now, and watch what Intel is able to bring. My thinking is that AMD never took seriously GPU market appart from console gaming which doesn't have to compete but only on the price. GPu at AMD was only an argument for selling CPU. Though it is their APU that made their sales until recently against Intel. Now they don't even need to bother about their APU, and Navi shows that.
Navi is a very poor upgrade from Vega. In fact it trades the compute for gaming when Turing adds all the specs together. It's still on 12nm now 11nm, and has better gaming, better computing (now better than AMD Vega) and brings hardware Raytracing AND IA Tensor cores. Nvidia is now lightyears ahead of AMD.

I must say, that everything you posted is in fact a fib.

If you go back to when AMD bought ATi, they had a goal, to be the first company to bring heterogenous computing to the masses. AMD's CEO (Dr Lisa Su) has made it a point that their next round of 7nm APU will be the realization of that dream. AMD has more advanced technologies than Nvidia, or even Intel because they have to be innovative over the last 10 years.

As such, I could not let your post go in good standing, being so frought with fibbery.



Secondly, in just a few days, AMD will be releasing their 2nd generation of 7nm GPUs (Radeon VII being the 1st Gen). And within 6 months from now, AMD will be releasing their 3rd generation 7nm GPU.

In contrast, I do not believe that Nvidia has even taped out their new 7nm GPU at Sumsung, so I think you are a tad over-zealous in when you will see Nvidia's next Gaming GPU. Perhaps, little over a year from now around the Nov 2020, next year.
 
Last edited:
I must say, that everything you posted is in fact a fib.

If you go back to when AMD bought ATi, they had a goal, to be the first company to bring heterogenous computing to the masses. AMD's CEO (Dr Lisa Su) has made it a point that their next round of 7nm APU will be the realization of that dream. AMD has more advanced technologies than Nvidia, or even Intel because they have to be innovative over the last 10 years.

As such, I could not let your post go in good standing, being so frought with fibbery.



Secondly, in just a few days, AMD will be releasing their 2nd generation of 7nm GPUs (Radeon VII being the 1st Gen). And within 6 months from now, AMD will be releasing their 3rd generation 7nm GPU.

In contrast, I do not believe that Nvidia has even taped out their new 7nm GPU at Sumsung, so I think you are a tad over-zealous in when you will see Nvidia's next Gaming GPU. Perhaps, little over a year from now around the Nov 2020, next year.

The problem for AMD is that they on 7nm cannot beat NVIDIA on 12 nm...that does not bode well.
 
I bet Nvidia shader based algo is definitely not the best what can be achieved without dedicated hardware

Yeah definitely not. But all you have to do is look at highly optimized shader based raytracing software like Optix to understand the vast difference in performance.

We’ve been doing RT in software for a very long time on both GPUs and CPUs. It’s unlikely though that nvidia’s software DXR implementation is as optimized as their Optix stuff.
 
Yeah definitely not. But all you have to do is look at highly optimized shader based raytracing software like Optix to understand the vast difference in performance.

We’ve been doing RT in software for a very long time on both GPUs and CPUs. It’s unlikely though that nvidia’s software DXR implementation is as optimized as their Optix stuff.
Software DXR is pretty much emulating exactly what RT cores do (and to some extent Tensor cores for denoising) so this alone could mean this is not the best solution for shader based ray-tracing. There are perhaps more clever ways this problem could be tackled that does not translate well into hardware by the rule: if clever optimizations cost more transistors than adding more slower optimized hardware then for things that can be parallelized just add more simpler hardware and brute force it. Ray tracing is very highly parallelizable by its nature. Whole GP-GPU idea works on this idea. Single CUDA core is pretty pathetic compared to even simplistic CPU core but sheer amount of them does the job.

It might be that AMD will surprise us all and show some clever software DXR implementation. Their shader capabilities when executing complex software were after all always pretty good and when combined with clever usage of other units it might be actually pretty good. Of course pretty good for software solution and I highly doubt it will rival RT cores, especially in accuracy (AMD might want to sacrifice accuracy for speed if all they have is SW solution. they are after all not obliged to have the same quality and in most cases it is not really needed anyway...) Thankfully soon we will find out so no more guessing will be needed :)
 
Back
Top