Big Navi VRAM specs leak: 16GB Navi 21 and 12GB Navi 22 go head to head with GeForce RTX 3080 and RTX 3090

+15% performance increase over a 2080 Ti is something we've been hearing for a while. I don't see that necessarily being a bad thing as that would make it only about 10-15% slower than a 3080. Give it 16GB of VRAM and price it at $600 and it'll be a solid contender against the 3080 and would also likely persuade some prospective 3070 buyers to spend an extra $100 for double the VRAM and ~10-20% more performance.

Actually that is what coretek sources told him a while back and he hasn't confirmed anything else from him, but others do have deeper sources. Which is probably an AIB. MLID (Moores law is dead) and few others have said the Skew that has been leaking is not the top die. AMD has the top die in house and no one has access to it outside of AMD. Been keeping it close to chest. But the leaks are not of the top die because AMD is keeping that in house.
 
He's probably thinking of Raja's claims that 2x 480s in Crossfire would equal a 1080.
Yeah, crossfire makes a bit more sense when it scales well, unfortunately that was a very limited to a handful of games and has sadly gone the way of the dodo bird.
 
Sadly ended at vega and yes scaling was goin well if implemented in game. Could run RDR2 at 4K with reduced shadows and some other non essentials reduced a bit on a pair of 64's. Only in winter tho lol!
 
RX 480 Crossfire was actually decent when it worked.

I wished I saved some of those benchmarks, I was getting around 80 fps with GTA V almost max at 4K.

The older games usually worked better, and there were some glitches, but performance was there.
 
Come on AMD... put out a 3090 equivalent GPU for $719.. you have a process advantage, where are the chips...?

Wishful thinking I suspect. But I am an optimist.
 
I'll take 3080 performance for $719 as long as I can actually buy the damn thing at launch!
 
Yeah demand for the 3xxx cards is nuts. People were waiting in line at Microcenters' 4 (might have been more) days in advance.

For me the AMD needs to be cheaper at the same performance for it to be a consideration. Their drivers eventually catch up but you shouldn't need to wait a year after a game launch to get the proper performance.
 
The supply situation with the 3080 has discouraged me from trying to get one for now so I've decided to wait for RDNA 2. The leaked specs and features such as the VRAM size and Infinite Cache are making RDNA 2 very intriguing to me. I can see it matching or beating the 3080 in rasterization performance but ray tracing performance will be the deciding factor in my purchasing decision.
 
Well I am just going to wait and see what AMD has to bring. Nvidia would of had my $1500 of they had stock.
 
CFA6CB87-2FA6-4769-BDDD-3A0846257123.png
 
That didn’t explain the significance of the graph. Why does Big Navi appear to be better in one and Ampere better in the other? Which of the two performance, ray-box or ray-triangle testers, is more important?

"Level 2. Ray-box and ray-triangle testers can be implemented in hardware using standard fused multiply-add operations on GPUs but this repeated operation is expensive (cycles/power/area cost). A Level 2 solution offloads a large part of the ray tracing job to dedicated hardware improving efficiency. "

I forgot how much everyone likes to be spoon fed around here. A deep dive into level 2 ray tracing is way beyond the scope of this forum or thread. Just take it as more evidence that Big Navi should be 3080 competitive from these metrics.
 
That didn’t explain the significance of the graph. Why does Big Navi appear to be better in one and Ampere better in the other? Which of the two performance, ray-box or ray-triangle testers, is more important?
Basically, they are two different methods for determining ray collisions on the surface of an object, each has its own set of pros and cons.
 
"Level 2. Ray-box and ray-triangle testers can be implemented in hardware using standard fused multiply-add operations on GPUs but this repeated operation is expensive (cycles/power/area cost). A Level 2 solution offloads a large part of the ray tracing job to dedicated hardware improving efficiency. "

I forgot how much everyone likes to be spoon fed around here. A deep dive into level 2 ray tracing is way beyond the scope of this forum or thread. Just take it as more evidence that Big Navi should be 3080 competitive from these metrics.
Ray box is in early out detector (aka, if you have 15 triangles in a box, and said ray doesn't intersect said box, no need to check triangles inside); ray triangle is for determining surface that it actually hit and whether it needs to do more bounces. Both are important, which is more important is going to be determined by the game engine and how it stores it's data. If every triangle is in its own box, then ray-box is more important... If the entire scene is one box full of triangles, then ray-triangle is more important. Also, things that aren't reflections, like ambient occlusion, tend to use boxes in 3d space, so ray-box would be more important for this. It's really hard to say without diving into each specific game and how they are using it though, so take the graph with a large grain of salt (as with any leak).
 
"Level 2. Ray-box and ray-triangle testers can be implemented in hardware using standard fused multiply-add operations on GPUs but this repeated operation is expensive (cycles/power/area cost). A Level 2 solution offloads a large part of the ray tracing job to dedicated hardware improving efficiency. "

I forgot how much everyone likes to be spoon fed around here. A deep dive into level 2 ray tracing is way beyond the scope of this forum or thread. Just take it as more evidence that Big Navi should be 3080 competitive from these metrics.
Literally that gave me no information on what the graph is showing. Get off your high horse. I could spend hours reading various articles or I can ask someone to give me a brief summary who is experienced in the subject matter as the person below did. Not everyone is a pretentious asshat.
Ray box is in early out detector (aka, if you have 15 triangles in a box, and said ray doesn't intersect said box, no need to check triangles inside); ray triangle is for determining surface that it actually hit and whether it needs to do more bounces. Both are important, which is more important is going to be determined by the game engine and how it stores it's data. If every triangle is in its own box, then ray-box is more important... If the entire scene is one box full of triangles, then ray-triangle is more important. Also, things that aren't reflections, like ambient occlusion, tend to use boxes in 3d space, so ray-box would be more important for this. It's really hard to say without diving into each specific game and how they are using it though, so take the graph with a large grain of salt (as with any leak).

Thank you this was actually helpful.
 
Last edited:
Literally that gave me no information on what the graph is showing. Get off your high horse. I could spend hours reading various articles or I can ask someone to give me a brief summary who is experienced in the subject matter as the person below did. Not everyone is a pretentious asshat.
Just because you didn't understand the article doesn't mean there is not enough information there for someone else to understand.

Thank you this was actually helpful.
You've learned nothing in reality - except that if you wait around long enough, someone will give you the answer you want.
 
Just because you didn't understand the article doesn't mean there is not enough information there for someone else to understand.


You've learned nothing in reality - except that if you wait around long enough, someone will give you the answer you want.
I left with a better understanding of the graph than your lackluster response did. If you don’t want to “spoon feed” people asking for clarification then just don’t respond. Better off than linking an article that provided zero explanation to what I was asking in the first place. This is a hardware forum with people of varying knowledge; heaven forbid productive discussion occurs where people more versed in a topic can help educate those that aren’t.
 
Just because you didn't understand the article doesn't mean there is not enough information there for someone else to understand.


You've learned nothing in reality - except that if you wait around long enough, someone will give you the answer you want.

Article? its a graph with chinese characters on top. Yes someone could google tri-rate and box-rate, but asking on a tech forum for some clarification or cliff notes is hardly something to be offended over.

Expecting everyone to have the same base knowledge is laughably stupid though.
 
I left with a better understanding of the graph than your lackluster response did. If you don’t want to “spoon feed” people asking for clarification then just don’t respond. Better off than linking an article that provided zero explanation to what I was asking in the first place. This is a hardware forum with people of varying knowledge; heaven forbid productive discussion occurs where people more versed in a topic can help educate those that aren’t.
Consider it done. Thank you for that education.
 
I don't care my point literally still stands that there is nothing wrong with asking for details on a tech form. plus with your pretentious first paragraph and little hissy fit in the second paragraph I am not inclined to click your link.

someone sure peed in your cheerios today.

You don't care? Then why reply? The linked article is neither in chinese or just a graph. I'm not referring to the original screenshot, so why are you?

Comprende?
 
Man, you'd think this man just asked to have your kidney, lol. He was just curious what the differences where. I didn't even read your article, but I have written my own raytracing software, so I understand the differences. Took < 2 minutes of my time to make a short summary of what I know that is easy to understand.

And.. after reading the link, I wasted more time digesting the information that barely even describes what each one is used for, just that they are implemented and it uses boxes for hierarchy, nor does it even mention how ambient occlusion typically works in bounded volumes (aka, boxes) in a scene in a 3d game, so reading it wouldn't have really been any closer to understanding what each bar actually represents that he could relate to. Aka, he'd of wasted 10+ minutes with less understanding than just giving a simple answer. If he wants to delve into the deep trenches, I'm not going to write a dissertation on raytracing, do any of his homework, or share my code, but a quick high level idea isn't that big of a deal.
 
I have written my own raytracing software, so I understand the differences.
I know you mentioned in a post above that both are important but is there one that is more important than the other? Will Ampere have the advantage in most cases with ray tracing or RDNA 2?
 
If above chart is correct.
In a nut shell.When a scene calls for G Box.Navi 21 will blaze thru it.Nvidia card will slow down.Nvidia appears to have upper hand in RT.

Reading AMD patents on this topic.Appears Big Navi or future AMD gpu will have a ability to redirect prioritys.

Let’s say Ampre has 20% of its hardware dedicated to RT. When RT is not needed it’s a waste of 20% rendering power.

Interesting read https://www.hardwaretimes.com/amd-big-navi-to-pack-80cus-5120-cores-20-cus-to-be-used-for-ray-tracing-20-tflops-of-fp32/amp/#referrer=https://www.google.com&amp_tf=From %1$s
 
I know you mentioned in a post above that both are important but is there one that is more important than the other? Will Ampere have the advantage in most cases with ray tracing or RDNA 2?
It really is implementation specific. I know that's an easy way out, but unfortunately it's true. A lot of game engines use bounding boxes of sorts (axis aligned bounding boxes) in which you stuff a bunch of triangles into a box. You have cubes in 3d space. So, take your entire level (this is just for a visual) and split it into 4 boxes. Now you have ~1/4 the triangles per box. Now take these boxes and split them all into 4, each how has about 1/16th the triangles, etc. If a box has a lot of triangles (and this is where it's implementation specific, there is no single # for a lot) you divide it further, if a box has few or no triangles, then you stop splitting it. This allows the sky that has no triangles to not take up much extra memory (And allows ray casting towards the sky to not have to intersect with many things). So, when we cast a ray through this world, we first are going to test all the ray-box intersections to see if a ray hits a box, if so, then we must go through ALL triangles in that box using ray-triangle intersections to find what it actually hit (or if it even hit anything at all!!). So, depending on how the hierarchy of the game is done it may favor one over the other, so without knowing the specific game or how it segments the world, I really can't say which is more important. Current implementations (not sure about all of them, as I don't have the privilege of seeing everyones implementations) of Ambient Occlusion tend to store data in a grid of boxes (a 3d texture that stores light information at different points of the area it covers), so for ambient occlusion (that's implemented in this manner) it is 100% (or close to it) ray-box testing (it's ever more complicated as they normally use cone tracing, where the size of the cone depends on the material specifics, high specular means it's mirror like, so narrow cone, high diffuse means matte which uses a larger cone). Of course, all of this is simplified and there are many other nuances involved ;).

Sorry I can't just give an A is better than B answer, as it's really not that simple. From my 1,000ft view, AMD will probably have better performance in AO and NVidia will do better with reflections (which makes me laugh that AMD's RT demo was almost all reflections), but really it will be determined by the way the game engine works internally and this guess/view can completely fall apart if optimized for one vs the other.
 
It really is implementation specific. I know that's an easy way out, but unfortunately it's true. A lot of game engines use bounding boxes of sorts (axis aligned bounding boxes) in which you stuff a bunch of triangles into a box. You have cubes in 3d space. So, take your entire level (this is just for a visual) and split it into 4 boxes. Now you have ~1/4 the triangles per box. Now take these boxes and split them all into 4, each how has about 1/16th the triangles, etc. If a box has a lot of triangles (and this is where it's implementation specific, there is no single # for a lot) you divide it further, if a box has few or no triangles, then you stop splitting it. This allows the sky that has no triangles to not take up much extra memory (And allows ray casting towards the sky to not have to intersect with many things). So, when we cast a ray through this world, we first are going to test all the ray-box intersections to see if a ray hits a box, if so, then we must go through ALL triangles in that box using ray-triangle intersections to find what it actually hit (or if it even hit anything at all!!). So, depending on how the hierarchy of the game is done it may favor one over the other, so without knowing the specific game or how it segments the world, I really can't say which is more important. Current implementations (not sure about all of them, as I don't have the privilege of seeing everyones implementations) of Ambient Occlusion tend to store data in a grid of boxes (a 3d texture that stores light information at different points of the area it covers), so for ambient occlusion (that's implemented in this manner) it is 100% (or close to it) ray-box testing (it's ever more complicated as they normally use cone tracing, where the size of the cone depends on the material specifics, high specular means it's mirror like, so narrow cone, high diffuse means matte which uses a larger cone). Of course, all of this is simplified and there are many other nuances involved ;).

Sorry I can't just give an A is better than B answer, as it's really not that simple. From my 1,000ft view, AMD will probably have better performance in AO and NVidia will do better with reflections (which makes me laugh that AMD's RT demo was almost all reflections), but really it will be determined by the way the game engine works internally and this guess/view can completely fall apart if optimized for one vs the other.
Thank you for the explanation! It's starting to make more sense. I guess my only worry is that we'll get some games that are crippled on one architecture and run fine on the other. Hopefully most games will have some sort of balance.
 
Thank you for the explanation! It's starting to make more sense. I guess my only worry is that we'll get some games that are crippled on one architecture and run fine on the other. Hopefully games will have some sort of balance.

I don't think its going to matter much, raster is still going to be the standard for years if not the next decade, ray tracing will be largely optional and hit n miss.
 
Thank you for the explanation! It's starting to make more sense. I guess my only worry is that we'll get some games that are crippled on one architecture and run fine on the other. Hopefully games will have some sort of balance.


From the looks of things, Both the Box and Tri are better than the OG Titan RTX, so I don't imagine it will be suffering in any current title.

The truth is that most (if not all) RT operations will user a mix of both BVH and Triangle intersection, some will lean on one more than the other.
 
Remember nvidias tessellation back in the day.How it ran or lack there of On ATI/AMD hardware of the time.

Its all about game code to bring out the strength or weakness in either green or red.
 
Remember nvidias tessellation back in the day.How it ran or lack there of On ATI/AMD hardware of the time.

Its all about game code to bring out the strength or weakness in either green or red.

I beleive the underlying software is more even these days (Dx12/vulkan) particularly for tech like ray tracing. This isn't gameworks.
 
Thank you for the explanation! It's starting to make more sense. I guess my only worry is that we'll get some games that are crippled on one architecture and run fine on the other. Hopefully most games will have some sort of balance.
Sure thing, hopefully some folks can gain a little insight/knowledge, or at least get a kick out of my wasted time :p.

Yeah, this is why games optimized for XBOX/PS5 *may* favor AMD, because they are the most limited and will be leaning towards less tri, more box (and this also makes things like collision detection cheaper for the physics engine), however there has to be a balance, as each of those bounding boxes takes up some memory. You can't just draw a box around every single triangle and expect it to work :). Then again, you can't just throw all triangles in a single box and expect it to work either. There is a balance that devs will have to find, and possibly in the future may even have some sort of tweaking that goes on to maximize performance of both platforms if they feel so inclined to code it up (which is possible, but that's a bit of effort for what amounts to very small niche group of people that care). Just based on past experience, I suspect nvidia will still have superior performance in RT in general due to more optimized code for the larger market share, although with consoles and piss poor job some companies do at porting from console to PC may give AMD some wins here and there. One things for sure, nothing is set in stone and we all really won't know (not even when RDNA2 drops, as games will still need to catch up) what kind of performance can be teased out for some time. Heck, game devs are still trying to figure out the best way to utilized nvidia's RTX when it's been out for a while (I personally like its use in lighting more than a million reflective puddles). Any emerging tech is going to take time to get implemented, then it has to be optimized and then people figure out the best ways to use it (aka, lighting, reflections, physics, 3d audio, all the above?). It's getting interesting, as the last few releases haven't really been that interesting (to me at least, RTX cards were semi-interesting, but I knew there wouldn't be much in the way of implementation for some time). Ampere seems to be a reasonable performance increase and RDNA2 seems to be coming together well, so I'm hopeful we'll have some much needed competition which will drive some more innovation. I'm really curious to see how much performance the RDNA2 boards run with limited memory bus width and a small onboard (on fabric?) cache. Right now to me, that's the biggest unknown as we know approximately what 80CU's *should* be able to do based on what 40CU's currently do (plus a few percent for architecture improvements), but what they actually produce will be interesting.
 
Remember nvidias tessellation back in the day.How it ran or lack there of On ATI/AMD hardware of the time.

Its all about game code to bring out the strength or weakness in either green or red.

Ah yes, tessellating things to such a degree that the resulting triangles are smaller than a pixel at the rendered resolution and getting a developer to place tessellated water that extends to the horizon under the map where nobody can see it just because you know your geometry performance is massively overspec'd and it will hurt the competition. Classic NVIDIA.
 
If above chart is correct.

Where’s the chart from? Was there any other info on how performance was tested?

Let’s say Ampre has 20% of its hardware dedicated to RT. When RT is not needed it’s a waste of 20% rendering power.

Yeah that’s the trade off with dedicated hardware. It’s wasted silicon when not used but is supposed to be much faster at the specific thing it’s optimized for. If AMD is able to pull off better performance with general purpose hardware they will have quite the coup.
 
Back
Top