Navi 21/23 Cards Rumored (aka "Nvidia Killers" xD)

[leadeater]

1 minute ago, RejZoR said:

Then why my GTX 1080Ti sucks at ray tracing? I got shitloads of compute units?!

Because it doesn't have dedicated BVH hardware.

[leadeater]

@RejZoR

To keep it extremely simple what you are saying is equivalent to 'addition can only be used to count apples'. I'm pretty sure you can think of more uses for addition than to count apples. So if I were to make a hardware accelerated addition logic to count apples you're sure as heck going to be able to count oranges with it, it doesn't matter if I labeled the hardware logic as 'The Apple Adder' you can still use the hardware to count oranges so long as I give you a means of access to that hardware.

[Maticks]

I seriously doubt Navi will be able to hit RTX 2080ti speeds, but if they do it will be what 1000 Watts?

When AMD hit the numbers they usually also hit the PSU as well.

 

It's good for competition so i do hope AMD gets there, i certainly want to upgrade from this GTX1080 at some point soon.

$2,100 AUD for an RTX 2080ti is just too much to justify. $1,200 AUD i deem top end and it was 1080ti pricing back then but this is just crazy pricing right now.

[Bananasplit_00]

11 hours ago, Results45 said:

Navi 23 (92 CUs)

i want this, i want this so bad, but i doubt they are making it and if it does come to market its going to be soooo expensive

[hobobobo]

All i want is for amd to make their own kaby lake g chip with hbm on board, goddamn that would be a sick APU. I think amd patented their own implementation of emib some time ago so 1 zen2 die, 40 navi CU and 16gb hbm2.... ah, what a dream. Makes me somewhat sad that there is no real market for that unless you sell it to chinese lol and dota maniacs

[RejZoR]

1 hour ago, leadeater said:

@RejZoR

To keep it extremely simple what you are saying is equivalent to 'addition can only be used to count apples'. I'm pretty sure you can think of more uses for addition than to count apples. So if I were to make a hardware accelerated addition logic to count apples you're sure as heck going to be able to count oranges with it, it doesn't matter if I labeled the hardware logic as 'The Apple Adder' you can still use the hardware to count oranges so long as I give you a means of access to that hardware.

No it's not. Just because you have the logic to count apples, that doesn't mean you can count apples fast enough to be usable. That's the whole point I'm making.

[cj09beira]

17 minutes ago, Maticks said:

I seriously doubt Navi will be able to hit RTX 2080ti speeds, but if they do it will be what 1000 Watts?

When AMD hit the numbers they usually also hit the PSU as well.

 

It's good for competition so i do hope AMD gets there, i certainly want to upgrade from this GTX1080 at some point soon.

$2,100 AUD for an RTX 2080ti is just too much to justify. $1,200 AUD i deem top end and it was 1080ti pricing back then but this is just crazy pricing right now.

amd just made a whole new thing its bound to have plenty of places for them to optimize, plus with how big those dies will be they could lower frequency a little bit 100-200mhz almost not loose performance and lower power a whole lot, and those should be using hbm which will help as well 

33 minutes ago, hobobobo said:

All i want is for amd to make their own kaby lake g chip with hbm on board, goddamn that would be a sick APU. I think amd patented their own implementation of emib some time ago so 1 zen2 die, 40 navi CU and 16gb hbm2.... ah, what a dream. Makes me somewhat sad that there is no real market for that unless you sell it to chinese lol and dota maniacs

there isn't because its not out yet, with people wanting to go small it should do quite well, and if the cpu has access to that hbm then gaming performance with dedicated graphics should be great 

[Trixanity]

1 hour ago, Maticks said:

I seriously doubt Navi will be able to hit RTX 2080ti speeds, but if they do it will be what 1000 Watts?

When AMD hit the numbers they usually also hit the PSU as well.

 

It's good for competition so i do hope AMD gets there, i certainly want to upgrade from this GTX1080 at some point soon.

$2,100 AUD for an RTX 2080ti is just too much to justify. $1,200 AUD i deem top end and it was 1080ti pricing back then but this is just crazy pricing right now.

That's not how it works. You add more compute units and clock them lower. It's certainly doable in the 250-275W TDP range. Likewise AMD needs to abandon their yield strategy of prioritizing yields over anything else meaning up the price and tighten up the voltage: you've got the recipe for a power efficient behemoth of a chip that's faster than a 2080 ti.

 

Not to mention AMD pulling out more tricks to increase efficiency. There's plenty opportunity if AMD gets their act together.

[GOTSpectrum]

3 hours ago, leadeater said:

Then remind me again how GPUs can only be used for rendering images and cannot be used for other purposes and don't exist in every modern super computing cluster in the last 10 years?

 

Any hardware, any, that is doing math i.e all of them can be used for anything that math applies to. No matter how small and focused, if it's only a single thing it can be used for as many purposes as you can think of. The limitation is only in the ways you can think of to utilize that math or function. Yes not everything is easily reuable or makes sense to, as with video decoders those are not doing just one thing, one math calculation and has specific rules and flows for it's usage, not everything is that.

 

Nvidia made these cores for a specific task, purpose designed only for this one task. You know what those are called? RT cores. These cores can be used for more than just Nvidia RTX games because the math they are doing is not limited to only that usage. Why can they be used for more than that? Because all that is being done is BVH calculations.

https://www.anandtech.com/show/13282/nvidia-turing-architecture-deep-dive/5

 

If AMD is to make an accelerated hardware path for Ray Tracing it is required that they make this ASIC be able to do BVH. So if RT cores can be used for more than just Ray Tracing for RTX games AMD equivalent can too because they are doing the same thing. The only thing that would prevent this is either AMD or Nvidia not exposing that hardware through an API for you to use closing it off.

 

RT cores = BVH calculators = Applicable to more than Ray Tracing = Can do more than Ray Tracing.

Talking about GPUs having other uses...

 

FOLDING MONTH, get signed up now people!

 

Prizes to be won, bragging rights to be earned!

 

 

[ThePD]

Even if AMD could make these GPU's happen, Nvidia would respond with the 3000 series.

[sof006]

Guess I'll believe it when it happens. I mean everyone had doubts about Ryzen and now look at that.

[Unixsystem]

5 hours ago, RejZoR said:

No it's not. Just because you have the logic to count apples, that doesn't mean you can count apples fast enough to be usable. That's the whole point I'm making.

Do you have any evidence that there is a more effective single purpose ASIC solution to RT than one that exclusively works with BVH? That's what they were specifically built for, and while I don't doubt that we will get newer an possibly different ways to perform RT calculations as the technology matures, currently the RT cores seem to be the best anyone has come up with. 

 

Is it fast enough for reasonable use in most games as decent framerates? No. But saying that RT cores aren't fast enough and therefore there is a better solution is like saying that the 2080ti isn't fast enough for 8k60 and therefore we need a new GPU paradigm for it. 

[leadeater]

8 hours ago, RejZoR said:

No it's not. Just because you have the logic to count apples, that doesn't mean you can count apples fast enough to be usable. That's the whole point I'm making.

You still do not understand, my example was that the hardware is the fastest possible to count apples, it was purposed for counting apples faster than anything else. I may have designed the hardware and the API to only accept a one INT and one BIT inputs and the bit value is added to the INT value and the result returned you can see how that can be used for more than just counting apples. The label 'The Apple Adder' doesn't make the function the hardware was designed to do only work for apples.

 

Nvidia RT cores are doing BVH calculations, that is what they do. The reason for why that is what they do is that is how you cull rays to find only the important ones that actually intersect with objects in the view port.  The hardware RT cores are specifically made to do those calculations quickly, quicker than Shader Cores.

 

What you are saying is that BVH can only be used for Ray Tracing, or the RT cores or hardware like it that can do BVH very quickly can only be used to do Ray Tracing and this is not correct.

 

If I have hardware that can count extremely fast I can use it to count anything extremely fast, it doesn't matter if the original purpose was to count apples you can still use the counting hardware to count something else, so long as you need to count not multiply.

 

As long as someone needs to do BVH you can use RT cores for that, it's not a Ray Tracing only function, it is used for Ray Tracing because it was found suitable for use in Ray Tracing.

 

 

Quote

I've got a handful of scientific computing apps that run on CUDA, and a big performance bottleneck for these codes has been geometry search (e.g. locating parts of large meshes that are in contact). I've been using a BVH implementation based on the blog posts by Tero Karras:

 

This has been a huge win for performance so far, and I was excited to see that the new Turing cards have dedicated hardware for BVH traversal, which is by far the most expensive part of the search process for me.

https://devtalk.nvidia.com/default/topic/1038619/cuda-programming-and-performance/api-for-bvh-traversal-on-turing-gpus/1

 

Here is an example of someone doing BVH and not for Ray Tracing, he wishes to use the dedicated hardware in Turing for this, the RT cores.

[leadeater]

1 hour ago, Waffles13 said:

Do you have any evidence that there is a more effective single purpose ASIC solution to RT than one that exclusively works with BVH? That's what they were specifically built for, and while I don't doubt that we will get newer an possibly different ways to perform RT calculations as the technology matures, currently the RT cores seem to be the best anyone has come up with. 

 

Is it fast enough for reasonable use in most games as decent framerates? No. But saying that RT cores aren't fast enough and therefore there is a better solution is like saying that the 2080ti isn't fast enough for 8k60 and therefore we need a new GPU paradigm for it. 

Even then that doesn't actually change the original point, whatever someone comes up with that ASIC will be doing some math function and that doesn't mean that math function can only be used for Ray Tracing. It's not like the field of mathematics is doing work to find better Ray Tracing methods, it's the other way around. People in the hardware and development industry are looking for techniques to make tasks easier, if you have 1 trillion rays and you know only 10% of them are of actually any use then you look for a way to find only that 10%. So if you made hardware designed to find that 10% of rays that doesn't actually mean it can only do that.

[Results45]

10 hours ago, Bananasplit_00 said:

i want this, i want this so bad, but i doubt they are making it and if it does come to market its going to be soooo expensive

If Lisa Su and her teams wants to steal the consumer hardware crowns from both Team Blue (as they have for the time being) and Team Green (even if just for Q1 2020) badly enough, they will ~ lest they bide their time and wait until next year when Intel joins the dedicated graphics ring. 

[Unixsystem]

1 hour ago, leadeater said:

Even then that doesn't actually change the original point, whatever someone comes up with that ASIC will be doing some math function and that doesn't mean that math function can only be used for Ray Tracing. It's not like the field of mathematics is doing work to find better Ray Tracing methods, it's the other way around. People in the hardware and development industry are looking for techniques to make tasks easier, if you have 1 trillion rays and you know only 10% of them are of actually any use then you look for a way to find only that 10%. So if you made hardware designed to find that 10% of rays that doesn't actually mean it can only do that.

I'm not familiar enough with the mathematic background of these sort of tasks to know if BVH was absolutely the only way to effectively cull rays, or if there were other ways being researched that could conceivably replace it. Mostly just to give maximum benefit of the doubt to @RejZoR's point about an RT core replacement.

 

I think broadly everyone can agree that RT cores are not fast enough to accomplish their intended goals outside of heavily constrained circumstances, but that doesn't mean that there's some magic bullet and they need to be replaced with something completely different. More that they just need to be further optimized, which I'd imagine we'll get quite a bit of in Turing's followup, and possibly in whatever AMD's solution winds up being called. 

[Results45]

Here's a handful of sources on Bounding Volume Hierarchy (BVH) that explain what it is and how it helps speed up real-time rendering:

 

#1: Generally speaking.......https://en.wikipedia.org/wiki/Bounding_volume_hierarchy

 

#2: How it's implemented as an acceleration algorithm/structure for 3D rendering:

• https://www.scratchapixel.com/lessons/advanced-rendering/introduction-acceleration-structure/bounding-volume
• http://raytracey.blogspot.com/2016/01/gpu-path-tracing-tutorial-3-take-your.html
• http://raytracey.blogspot.com/2016/09/gpu-path-tracing-tutorial-4-optimised.html

 

#3: Volume Tiled Forward Shading by Jeremiah van Oosten (as in using BVH to optimize T.F. shading ~ especially with lots of lights sources):

• https://www.3dgep.com/volume-tiled-forward-shading/

 

#4: How Nvidia uses BVHs for ray-tracing in Turing/RTX: https://www.anandtech.com/show/13282/nvidia-turing-architecture-deep-dive/3

[leadeater]

2 hours ago, Waffles13 said:

Mostly just to give maximum benefit of the doubt to @RejZoR's point about an RT core replacement.

I was never doubting the point that it may be possible to make a hardware ASIC that is faster or uses a different mathematical method but rather that it could only ever be used for Ray Tracing. While there are ASICs that are more complex and have their own pipeline of work making their use outside of the intended purpose harder or not possible that's not all of them. SSE/AVX is an ASIC, Tensor Cores are an ASIC, RT Cores are an ASIC but it's possible to use these to achieve multiple different tasks.

 

I was only ever pointing out the mistake in how a fixed function piece of hardware is thought of, it's not a Fixed Function Ray Tracer its a Fixed Function [something here] used for Ray Tracing, and depending on what that something is and the constraints of it's usage could be used for tasks other than Ray Tracing so long as the tasks aligns with the constraints.

 

It's more important to know what the ASIC is actually doing than to look at the name given to it by the manufacture, if you know what it is doing then you can think of ways to use it and that doesn't have to be how it was intended to be used.

[Maticks]

It is Gen 1 Ray Tracing its not going to get any awards here, name a gen one product in tech thats delivered a 100 percent experience.

It's 40FPS with Ray Tracing On, the real value is going to be the 3080ti cards thats going to hopefully have 60+ FPS with Ray Tracing On.

 

[GamerDude]

13 hours ago, Bananasplit_00 said:

i want this, i want this so bad, but i doubt they are making it and if it does come to market its going to be soooo expensive

And the RTX2080 Ti isn't? I'm waiting for NAVI 23 myself, to complement my 3900X.....actually, I'm waiting for PowerColor NAVI 23 Red Devil (5900XT Red Devil if that's what it'd end up being called).  I seriously doubt that it'd cost more than an RTX2080 Ti (or an RTX3080 Ti) even if it were to incorporate HBM2E which isn't cheap.

[thorhammerz]

16 minutes ago, Maticks said:

the real value is going to be the 3080ti cards thats going to hopefully have 60+ FPS with Ray Tracing On.

I'd argue 4070 (or more likely, the 5070) when RTX gets sufficient developer support to actually look nice, in conjunction with hardware-accelerated RT being accessible at below the $600 price point.

[illegalwater]

21 minutes ago, Maticks said:

It is Gen 1 Ray Tracing its not going to get any awards here, name a gen one product in tech thats delivered a 100 percent experience.

It's 40FPS with Ray Tracing On, the real value is going to be the 3080ti cards thats going to hopefully have 60+ FPS with Ray Tracing On.

 

When is the last time you've looked at benchmarks? The RTX cards are already doing 60+ FPS with ray tracing on..

[Suika]

I'm extremely curious if the CUs in Navi can actually scale that well without becoming another monster like the R9 295X2. Like I don't know much about the actual silicon and architecture but I can't imagine what sounds like more than doubling the core to be less than 400W.

[Maticks]

41 minutes ago, System32.exe said:

When is the last time you've looked at benchmarks? The RTX cards are already doing 60+ FPS with ray tracing on..

ok slightly better.. but still no 100FPS at 1440p

[Energycore]

I just hope these GPUs aren't massive low yield chips that leave AMD selling at a loss. Wasn't Navi supposed to bring "scalability"? If these guys are multiple chip solutions then I'm on board.