• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Report: AMD expects “Big Navi” to compete with 3070, not 3080

CrustyBritches

Gold Member
Leaving this here


I think he's making some valid assumptions. Hopefully we will see an October release for RDNA 2 because I'm not willing to wait and risk not having a card in time for Cyberpunk. 1st-world issues, for sure, but many people share this sentiment.

Big Navi will probably slot right at, or slightly below a 3080 at $649 with 16GB VRAM. This might force Nvidia to drop $50 or introduce higher VRAM capacity models, then AMD will drop to $599 and tweet "jebaited". Nvidia might be bundling WD Legion and 1yr. GeForce Now with 3070 and above, so AMD should do 1yr. Game Pass PC with some tradionally AMD-oriented title like DIRT 5.
 

VFXVeteran

Banned
I think he's making some valid assumptions. Hopefully we will see an October release for RDNA 2 because I'm not willing to wait and risk not having a card in time for Cyberpunk. 1st-world issues, for sure, but many people share this sentiment.

One of the main things that has steered me away from AMD cards is their drivers. They can not seem to make good drivers for the gamut of games. And the support on the visualization side is lacking too (OpenGL, Vulkan, etc..). It's really not a card I would consider even if it could match Nvidia cards.

Big Navi will probably slot right at, or slightly below a 3080 at $649 with 16GB VRAM. This might force Nvidia to drop $50 or introduce higher VRAM capacity models, then AMD will drop to $599 and tweet "jebaited". Nvidia might be bundling WD Legion and 1yr. GeForce Now with 3070 and above, so AMD should do 1yr. Game Pass PC with some tradionally AMD-oriented title like DIRT 5.

Unfortunately, this is not going to be a black and white thing. AMD will have to compete against a variety of modes when running games and I think they will definitely lose with Tensor + RT
 
Last edited:

Ascend

Member
And the support on the visualization side is lacking too (OpenGL, Vulkan, etc..)
I'm surprised you put Vulkan in there, considering that they practically made Vulkan. Technically it was Mantle, but, Mantle was donated to Khronos, which rebranded it as Vulkan.
 

VFXVeteran

Banned
I'm surprised you put Vulkan in there, considering that they practically made Vulkan. Technically it was Mantle, but, Mantle was donated to Khronos, which rebranded it as Vulkan.

Extensions. I haven't seen a RT demo on an AMD card yet. They are that far behind.
 

VFXVeteran

Banned


Or are you referring specifically to Vulkan? Is there a reason why it would work for DX12 and not Vulkan?


Vulkan. Yea, two different APIs with different function calls and different drivers.

Anymore demos other than this one with reflections?
 
Last edited:

VFXVeteran

Banned
Just that one, as far as I know. Unless you want to count the Xbox Series X;



Yea, but no. That's telling right there. SM units are used for RT and that will decrease shader performance especially since they are using them concurrently. I expect to see significant performance drops using all forms of RT. 4k is out of the question.
 

nochance

Banned
Leaving this here


Based on the XSX chip, which has hardware BVH acceleration, we know that 56 CUs take up around 300 mm^2 of space. Going from 56 to 80 is an increase of 43%, so crudely scaling everything by 1.43 that would be 429 mm^2 , which is already below the estimated 505 mm^2. Based on the Minecraft demo, it seems the XSX GPU is about on par with a 2070 for raytracing. Add another 54% (52 active to 80 CUs) and you are at 2080 Ti levels, not accounting for clockspeed improvements. So it makes sense to think that Navi 21 could be between the 3070 and 3080 for ray tracing, especially in "hybrid" titles that rely on conventional rasterisation.

Edit: Just to add, I think you are confusing silicon failures, which result in chips that have to be cut down or thrown away, and product failures caused by faulty components. A chip of 500 mm^2 would be perfectly in line with past AMD chips. And if 500 mm^2 is big, then what is the ~627 mm die Nvidia is using for the 3080?
If you mean the Minecraft tech demo that was shown to Digital Foundry, how did you manage to take any conclusions out of that? We know that for 2070 the average framerate is around 50 fps at full hd - what is the framerate on XSX? And is the demo rendering the exact same thing that the game does on a pc?

The next jump (albeit based on wild, baseless speculation) would make sense. But the last one, another 35-50% which is yet again absolutely massive, would stem simply from clockspeed improvements?

And what about more complex games that require compute units to do actual scene rendering in a traditional manner? How will that work if we already assumed that the entire die is going to be used for compute units - no ray tracing at all?

As for die size, we know, that AMD was experiencing problems with smaller dies, bigger die and a more complex solution naturally means more risk of failure. You can't just scale the die up by gluing extra compute units on, you need to take things like redundancy and heat aggregation into account. Somehow, law of physics no longer apply to AMD.

AMD hype trains are entertaining to watch, but ultimately not healthy for the fanbase. It is AMDs fault, as they keep overpromising and using misleading marketing.
 
Yea, but no. That's telling right there. SM units are used for RT and that will decrease shader performance especially since they are using them concurrently. I expect to see significant performance drops using all forms of RT. 4k is out of the question.
Does AMD even have dedicated blocks for RT like Nvidia with the RT cores? Or are they just using the compute cores for RT? Because if they are...that’s basically Pascal. And we all know how well Pascal does RT compared to Turing, to say nothing of Ampere.

I already know that AMD doesn’t have Tensor cores, which rules out any kind of AI/ML-based upsampling like DLSS. But if they don’t even have dedicated RT cores, it’s going to be an absolute disaster. Running RT solely on the compute cores will absolutely murder throughput on any scene which has both ray tracing and shaders, which is to say, all scenes that RDNA2 will try to render.
 

VFXVeteran

Banned
Does AMD even have dedicated blocks for RT like Nvidia with the RT cores? Or are they just using the compute cores for RT? Because if they are...that’s basically Pascal. And we all know how well Pascal does RT compared to Turing, to say nothing of Ampere.

I already know that AMD doesn’t have Tensor cores, which rules out any kind of AI/ML-based upsampling like DLSS. But if they don’t even have dedicated RT cores, it’s going to be an absolute disaster. Running RT solely on the compute cores will absolutely murder throughput on any scene which has both ray tracing and shaders, which is to say, all scenes that RDNA2 will try to render.

From all the rumors, it seems like there are no specific RTX cores like Nvidia.
 

Myths

Member
3070 was tempting but I cant in good conscience pay more than £400 for a card with only 8gigs. I really hope AMD aren't complete cunts and do the same.

I'll never in a million years spend £600 for a card, so 3080 was never in the running for me anyway.

Used 2080ti tanking in price are gonna be difficult to ignore.
Let me know. 2080TI’s are going to be jettisoned from PC cases like flyers from a plane. Worse for those who invested just before the announcement.
 
No, although much faster than 3070 for 500 is likely (and with 16GB).
3070-ish perf (up to 10% slower) for 400


5700 series rocked (trounced 2060/2070 and even upset 2080, forcing NV to react).
A 250mm2 chip.

Big Navi should be either 485mm2 or 505mm2. Significantly smaller than A102 (627mm2) but on a better process.


It would help if gamers would use own eyes and, god forbid, brains, instead of trusting that curious DF Truppe, that "for some reason" got that exclusive preview and claimed that if you go far enough from your screen, you could not notice the difference between 720p and 4k. It's a bloody upscaling, with all the shit that comes with it, like, you know, artifacts that DF "for some reason" did not see.


Was 10%-ish behind. In that sense, so will the next gen.
It's just, people don't click as eagerly if drama levels are low.


Dude, where does "tensor cores are used for ray tracing" come from pretty please?


What on earth are you on about? Even as is, 5700XT at 350 Euro is not too bad vs 3070 (which is 30%-ish faster)
Just curious, where do you live if you can buy the 5700XT for 350 Euros? Here in Sweden, they cost ~430-480 Euros.
 

Zathalus

Member
What the hell? How do they expect to have remotely competitive RT performance using just the TMU's + compute cores?
Xbox-DXR2.jpg

Xbox-SIMD.jpg

As you can see, the TMU units have been modified to accelerate BVH calculations. 3-10 fold increase in performance is in line with Nvidias own numbers.
 

Nydus

Gold Member
If they would announce their damn cards before September 17th they might have a chance to get a sale. But since we are talking Radeon group here they will sit on their asses till the end of September when everyone and their dog already bought Nvidia. I already sold my 2080Ti so whoever markets the better product to me faster, wins. But till now AMD released nothing to steal Nvidias Thunder.
 
If they would announce their damn cards before September 17th they might have a chance to get a sale. But since we are talking Radeon group here they will sit on their asses till the end of September when everyone and their dog already bought Nvidia. I already sold my 2080Ti so whoever markets the better product to me faster, wins. But till now AMD released nothing to steal Nvidias Thunder.
I honestly don't thick they have much to fight with Nvidia at this point. I feel like their raytracing will be behind Nvidia, not much to compete with DLSS, etc. I feel like they will have beasts in rasterization though. And good price to performance. But we shall see.
 

ZywyPL

Banned
Also, no DLSS on Big Navi is going to make things very rough in games which have DLSS enabled but that’s another story for another day

That rumored 3070 performance level is with DLSS off I assume? So when it's on the big Navi will have 3050 performance?


20% market share is kinda sad considering how long AMD/ATI has been in the GPU market.

Well it's also the result of how long AMD has been sitting in that awful GCN architecture.
 
Leaving this here



This guy is more accurate than most.

What does surprise me is he is saying there is even an answer to the 3090, although likely not quite as fast. I was only expecting a 3080 competitor. Which flies in the face of the OPs extremely tenuous rumours of 2080 Ti level performance snd that's it.

So the rumours of an 80 CU monster are likely true. Bring it on next month.
 
Last edited:

BluRayHiDef

Banned
This guy is more accurate than most.

What does surprise me is he is saying there is even an answer to the 3090, although likely not quite as fast. I was only expecting a 3080 competitor. Which flies in the face of the OPs extremely tenuous rumours of 2080 Ti level performance snd that's it.

So the rumours of an 80 CU monster are likely true. Bring it on next month.
I'm going with Nvidia regardless of what AMD offers, because Nvidia has better software support and better software features, and I've become accustomed to Nvidia's ecosystem.
 

Silver Wattle

Gold Member
That rumored 3070 performance level is with DLSS off I assume? So when it's on the big Navi will have 3050 performance?




Well it's also the result of how long AMD has been sitting in that awful GCN architecture.
AMD had been sitting in GCN so long because they were in a bad situation financially and then dedicated most resources to the Zen project, putting their GPU division in stagnation mode, AMD GPU division will be back soon enough now they have a much better fiscal position.
 
I'm going with Nvidia regardless of what AMD offers, because Nvidia has better software support and better software features, and I've become accustomed to Nvidia's ecosystem.

Sure but supply of RTX 3000 is extremely limited until AMD releases and probably all of this year anyway. So it wouldn't be sensible to try to get a 3080 this month, as you'll be paying over the odds. Perhaps if you live in the US it might be different. But the small quantities they have will be snapped up instantly. Samsung's yields are poor hence limited stock.
 
Last edited:

BluRayHiDef

Banned
Sure but supply of RTX 3000 is extremely limited until AMD releases and probably all of this year anyway. So it wouldn't be sensible to try to get a 3080 this month, as you'll be paying over the odds. Perhaps if you live in the US it might be different. But the small quantities they have will be snapped up instantly. Samsung's yields are poor hence limited stock.
I'll still try; I live near a Microcenter.
 

FireFly

Member
If you mean the Minecraft tech demo that was shown to Digital Foundry, how did you manage to take any conclusions out of that? We know that for 2070 the average framerate is around 50 fps at full hd - what is the framerate on XSX? And is the demo rendering the exact same thing that the game does on a pc?
According to Microsoft, the demo ran between 30 FPS - 60 FPS, so dead in the middle would be 45 FPS, which is the average for the RTX 2070 (https://www.tomshardware.com/uk/features/minecraft-rtx-performance). We don't know if the workloads are exactly equivalent, and we also don't know if the XSX can get faster with more shader optimisation. As for example, denoising is a huge performance sink, and is performed on the shaders even in Nvidia hardware. (One of the big reasons Nvidia doubled up FP32 compute was to improve denoising performance)

So to be clear, this is just informed speculation. I am not claiming we know XSX is on par with the 2070 in raytracing; just that it is within the realms of possibility.

The next jump (albeit based on wild, baseless speculation) would make sense. But the last one, another 35-50% which is yet again absolutely massive, would stem simply from clockspeed improvements?
Well, it's based on the consistent rumours of Navi 21 being a 80 CU part. I am not saying we should expect a 35% to 50% clockspeed improvement. I am saying that if the XSX performs around 2070 levels for raytracing, then a 54% increase in CUs would put it at around 2080 Ti levels, where the 3070 is, according to Nvidia. And then it would be reasonable to think AMD could increase clocks by 15% or 20%.

In that situation it would still be behind the 3080 in raytracing, but ahead of the 3070. The 3070 should be doing much better than a 2080 Ti, but it looks like it is being starved by memory bandwidth as it has 17% less than the 2080 Ti, and raytracing demands huge amounts of bandwidth.

And what about more complex games that require compute units to do actual scene rendering in a traditional manner? How will that work if we already assumed that the entire die is going to be used for compute units - no ray tracing at all?
In RDNA 2, each CU contains a raytracing unit that accelerates BVH calculations, just like with Nvidia. We know from the Hot Chips presentation that this only adds a small amount of extra die space, which explains why AMD can fit 56 CUs in ~300mm^2.

As for die size, we know, that AMD was experiencing problems with smaller dies, bigger die and a more complex solution naturally means more risk of failure. You can't just scale the die up by gluing extra compute units on, you need to take things like redundancy and heat aggregation into account. Somehow, law of physics no longer apply to AMD.
AMD has been using huge dies for years. Vega 64 was 486mm^2, Fury X was 596mm^2, R9 290 was 438 mm^2. Obviously cost, heat and power consumption scale with die size, which is why AMD had to improve performance per watt by 50% with RDNA2 and why AMD will be charging more for Big Navi! But in terms of their historical "big" GPUs, 500 mm^2 is a perfectly normal die size for AMD. I find it funny that you talk about AMD breaking the laws of physics when the 3080 is 627mm^2 and the 2080 Ti was 754 mm^2!

As far as failure rates go, again card failures != silicon failures. But if you are referring to the below article, then Asrock, MSI and Sapphire all had failure rates of 2%, which is equivalent to the best AIBs on the Nvidia side. Essentially it seems that Powercolor boards were not as well built, and brought the average failure rates up.

 
Last edited:

VFXVeteran

Banned
Xbox-DXR2.jpg

Xbox-SIMD.jpg

As you can see, the TMU units have been modified to accelerate BVH calculations. 3-10 fold increase in performance is in line with Nvidias own numbers.

This issue is that they are using TMUs and not separate cores. We won't know what bandwidth is taken from rasterization code until it runs in a game.
 

VFXVeteran

Banned
That demo was not just reflections, but RT AO and RT shadows as well:
Slide2.jpg

I'm convinced that AMD tacked on RT into some already wired TMUs to show that they have "hardware raytracing" but they really do not have them in the sense that the hardware was designed from the ground up to have separate RT cores. If it was performant, they would have been shown a game or something using all the forms of DXR. As of now, we only see reflections.
 
Last edited:

Nydus

Gold Member
We have to see if RT is really needed this gen. If developers don't use it, then it's the same all over again like this gen. DLSS is nice but again, is needed to be implemented by the developers. Maybe AMD being behind in RT won't matter as much in the next generation.
 

FireFly

Member
I'm convinced that AMD tacked on RT into some already wired TMUs to show that they have "hardware raytracing" but they really do not have them in the sense that the hardware was designed from the ground up to have separate RT cores. If it was performant, they would have been shown a game or something using all the forms of DXR. As of now, we only see reflections.
At this stage we are veering close to FUD.

There have been no demos of Big Navi, because the card hasn't been shown yet. However Microsoft has shown the XSX running Minecraft at 1080p at between 30 - 60 FPS.

The demo above was a DirectX demonstration, not a product demonstration. And the slide literally tells us that the shadow underneath the console the robot is using was generated via raytracing.
 

VFXVeteran

Banned
Well, the Microsoft presentation clearly states otherwise.

You can watch for yourself:


You can't show any of AO and GI when every material is a mirror. That's dumb. I can't see shit and notice they didn't focus on it at all. That's a DXR demo. Why isn't Nvidia running that demo? In short, it's a smokescreen. I don't see shit from it other than reflections. AMD can say they support RT and I believe them. It's just not in the sense of how Nvidia designed their boards. Also, they should have shown something concrete by now even if it's not finished. As of yet, we are all waiting on the AMD crew to come up with something meaningful. You have to admit that the information is sparse and not very persuading.
 
Last edited:

VFXVeteran

Banned
AMD's TMU's are extremely powerful though.

Yea, they are. Always had good TMUs. But I'm not going to assume it will have enough bandwidth to beat Ampere in RT/Rasterization combined tests. They aren't showing anything and we are 2 weeks from Nvidia release. That's just not confidence at all.
 
Last edited:

Zathalus

Member
You can't show any of AO and GI when every material is a mirror. That's dumb. I can't see shit and notice they didn't focus on it at all. That's a DXR demo. Why isn't Nvidia running that demo? In short, it's a smokescreen. I don't see shit from it other than reflections. AMD can say they support RT and I believe them. It's just not in the sense of how Nvidia designed their boards. Also, they should have shown something concrete by now even if it's not finished. As of yet, we are all waiting on the AMD crew to come up with something meaningful. You have to admit that the information is sparse and not very persuading.

Are you seriously telling me you cannot see the RT shadows in that demo? As for the AO, if you cannot notice it does not mean its not there. The is a entire presentation slide discussing it. So that demo is evidence of RT reflections, RT shadows, and RT AO all being done in the same scene.

As for showing something concrete, how does not this demo not qualify? That being said, the cards haven't been announced yet, but we have multiple RT examples in PS5 game announcements, and that is only a 10 TFLOP GPU, big Navi is apparently going to be more then double that.
 
This issue is that they are using TMUs and not separate cores. We won't know what bandwidth is taken from rasterization code until it runs in a game.
Yes, from what I have gathered over on the AMD subreddit, these are called "double-sided TMU's" because they can switch between running raster ops and running RT ops. The issue is of course if you do this, you can either do raster or RT but not both at the same time. So everytime the RDNA2 starts doing RT it's going to cut raster performance by however much RT is going into the scene.

Nvidia can do raster + RT at the same time, in addition to raster + compute and raster + RT + compute + tensor (DLSS). This RDNA2 design seems to imply it can do raster + compute, and RT + compute, but not raster + RT and last I checked you still need to render the scene at some point in your pipeline. So I'm baffled by this design. RDNA2 might actually have very competitive traditional rasterization performance, but the moment you turn on any kind of RT, you're going to tank raster perf and you'll also have to limit the amount of RT you do unless you plan on doing zero raster and that's uh not possible.
 
Last edited:

VFXVeteran

Banned
Are you seriously telling me you cannot see the RT shadows in that demo? As for the AO, if you cannot notice it does not mean its not there. The is a entire presentation slide discussing it. So that demo is evidence of RT reflections, RT shadows, and RT AO all being done in the same scene.

As for showing something concrete, how does not this demo not qualify? That being said, the cards haven't been announced yet, but we have multiple RT examples in PS5 game announcements, and that is only a 10 TFLOP GPU, big Navi is apparently going to be more then double that.

Ok. Maybe I'm not making myself clear.

Do you think that is a good representative demo of all of the RT features? Compare that demo to the Marble demo

 

VFXVeteran

Banned
Yes, from what I have gathered over on the AMD subreddit, these are called "double-sided TMU's" because they can switch between running raster ops and running RT ops. The issue is of course if you do this, you can either do raster or RT but not both at the same time. So everytime the RDNA2 starts doing RT it's going to cut raster performance by however much RT is going into the scene.

Nvidia can do raster + RT at the same time, in addition to raster + compute and raster + RT + compute + tensor (DLSS). This RDNA2 design seems to imply it can do raster + compute, and RT + compute, but not raster + RT and last I checked you still need to render the scene at some point in your pipeline. So I'm baffled by this design. RDNA2 might actually have very competitive traditional rasterization performance, but the moment you turn on any kind of RT, you're going to tank raster perf and you'll also have to limit the amount of RT you do unless you plan on doing zero raster and that's uh not possible.

Exactly. I'm not sure where the defensiveness is coming from here. What AMD has shown is very lackluster compared to what we know about RT.
 
Last edited:

Zathalus

Member
Ok. Maybe I'm not making myself clear.

Do you think that is a good representative demo of all of the RT features? Compare that demo to the Marble demo


No, its frankly a terrible demo in terms of looks. But we are not discussing the artistic merit involved, just what the demo shows regarding the technical capabilities RDNA 2 has. The demo clearly shows that RDNA 2 has the ability to render a huge number of reflections, combined with RT AO, and RT shadows.

If you need any further evidence of the merits of RDNA 2 RT performance, just look to the Minecraft demo on the XSX, it appears that the performance is in the same ballpark as the 2080, indicating that RDNA 2 at least matches Turing in RT capabilities. That being said, I doubt it will match Ampere in this regard.

I fully expect to see far more demos on RDNA 2 RT performance when they, you know, actually launch the cards.
 

FireFly

Member
Yes, from what I have gathered over on the AMD subreddit, these are called "double-sided TMU's" because they can switch between running raster ops and running RT ops. The issue is of course if you do this, you can either do raster or RT but not both at the same time. So everytime the RDNA2 starts doing RT it's going to cut raster performance by however much RT is going into the scene.
Nvidia can do raster + RT at the same time, in addition to raster + compute and raster + RT + compute + tensor (DLSS). This RDNA2 design seems to imply it can do raster + compute, and RT + compute, but not raster + RT and last I checked you still need to render the scene at some point in your pipeline. So I'm baffled by this design. RDNA2 might actually have very competitive traditional rasterization performance, but the moment you turn on any kind of RT, you're going to tank raster perf and you'll also have to limit the amount of RT you do unless you plan on doing zero raster and that's uh not possible.
My understanding was that texture filtering occurred at the end of the graphics pipeline, after geometry and lighting. So the ray intersection tests would be performed while the TMU would otherwise be idle.

But I am sure our resident expert VFXVeteran, can enlighten us.

(It's interesting that Ampere now has half the number of dedicated texture units per CUDA core).
 
Last edited:

Nydus

Gold Member
Here's what a real leaker had to say....A guy who is a virtual source...


Are there any leaks when the new cards will be revealed? I mean they surely can't wait till October. Even if they launch end of October it would be nice to know if waiting would be wise. :/
 
My understanding was that texture filtering occurred at the end of the graphics pipeline, after geometry and lighting. So the ray intersection tests would be performed while the TMU would otherwise be idle.

But I am sure our resident expert VFXVeteran, can enlighten us.

(It's interesting that Ampere now has half the number of dedicated texture units per CUDA core).
I thought the whole goal of modern GPU's was to keep occupancy as close to 100% as possible. I would imagine at a bare minimum, the TMU will already be active working on the next rendered frame because you know double or triple buffering, it won't just be idle because the current frame is done.
 

Ascend

Member
Exactly. I'm not sure where the defensiveness is coming from here. What AMD has shown is very lackluster compared to what we know about RT.
Wouldn't a bunch of mirrors be a lot harder to render than a bunch of barely reflective objects?
 
Top Bottom