Primitive Discarding in Vega: Mike Mantor Interview
HTML-код
- Опубликовано: 5 сен 2024
- Talking primitive shaders in Vega, small primitive discarding, and more with Mike Mantor of AMD.
Ad: Thermaltake Floe RGB (goo.gl/s88rjK)
Article: www.gamersnexus...
We have a new GN store: gamersnexus.sq...
Like our content? Please consider becoming our Patron to support us: / gamersnexus
** Please like, comment, and subscribe for more! **
Follow us in these locations for more gaming and hardware updates:
t: / gamersnexus
f: / gamersnexus
w: www.gamersnexus...
Links to Amazon and Newegg are typically monetized on our channel (affiliate links) and may return a commission of sales to us from the retailer. This is unrelated to the product manufacturer. Any advertisements or sponsorships are disclosed within the video ("this video is brought to you by") and above the fold in the description. We do not ever produce paid content or "sponsored content" (meaning that the content is our idea and is not funded externally aside from whatever ad placement is in the beginning) and we do not ever charge manufacturers for coverage.
GN's Andrew Coleman used Unreal Engine & 3D software to do a lot of the showcases/demos in this video! Definitely helps to visualize it (thumbs up for Andrew!), as some of these concepts are abstract. Partial transcript here, just for folks who want to be able to skim critical parts: www.gamersnexus.net/guides/3010-primitive-discarding-in-vega-with-mike-mantor
You might also be interested in our RX Vega final specifications coverage: ruclips.net/video/51SsMsUXTt8/видео.html
Love the interviews
Now that's some good interesting insightful BTS stuff
I was wondering where that showcase came from. Thanks Andrew!
a good question would have been if this tech is already working on vega FE, or how much of a performance improvement can we expect in real life applications from enabling this
Most of this is already in use and has been for a very long time. Some specific stuff, like DSBR, is not in use in FE but will enable for RX.
This is where other tech tubers (Like LTT) fall short.
Excellent stuff, Steve!
I strongly believe that this is the direction others should follow! This, and the comprehensive testing, that pretty much nobody else is doing!
LTT is catered to a completely different market. It's for the mass, while GN more or less targets enthusiasts.
@Patrick Tritp, exactly, Apples to Oranges (LTT Colored).
While I agree it is different, it is not "completely''.
The fact that both you and me(and whoever else) know of both GN and LTT, proves it.
mstoynov90 LTT is entertainment. GN is news. They fill different niches.
@mstoynov90
Um... Google recommendation?
AMD clearly has a lot of very smart people working for them. I often wonder if they get frustrated by the lack of R&D capital. I personally feel AMD could blow some minds if they had the money to spend. That said, I hear they are increasing their R&D budget, so maybe they're headed in a better direction now.
Amd just doesn't have enough market share to truly innovate. Look at vega, there's not much reason go for vega over a 1080, even including a slightly lower price.
I love this channel for Videos like this
I appreciate that instead of cutting out technical bits of the interview you add footage to seek to visually explain it. I think it's better to challange the viewer to learn more than try and dumb down the content.
I absolutely love this sort of in depth coverage of specific architectural information!
We have been knowing about this for over half a year, nice that they finally speak about it. The big question: overhead vs. efficiency gain due to less rendering.
This is one of the most interesting videos you guys have put out. Keep 'em coming!
With culling, what about light effects coming from objects not visible themselves?
Good question- perhaps a geometry reference is preserved?
TheEVEInspiration I think lighting references remain untouched even if the geometry is discarded. That way, if you get a world view and compare it to the player's view, the lighting comes from apparently empty space.
That depends on the game engine, the particular logic the developer assigned to the light, and how deeply it involves the CPU in these affects.
Assuming the light entity is given the go-ahead to apply post-processing to the rest of the scene, it is obviously post-processing so it happens *after* culling. (Meaning no shading occurs on unseen items because they have already been culled)
And as I said, the game engine decides whether or not to cull those.
Light and object are two different things. In games you have an object (say, a lamp) and a light source that positioned to match the position of the actual light source of the object you want to draw (where the bulb is). Basically, object is a physical thing on the screen that has vertices whereas light source is an abstract thing that only describes the light - its position, type of the source (point, directional etc), color, direction and other stuff. When you're culling you're only discarding vertices of the object. The light source that's attached to it in the game world is still there and later will be used to do the lighting. You don't need the object to do that (the lamp). In fact, in games there're many cases where you want to put light sources that has no object at all. For example, in cutscenes to better light character faces - you don't need a lamp for that, only the light source. Of course, that's very basic explanation and exact details could be very different depending on the graphics engine.
I think he means like what happens to objects out of scene that are casting shadows into the scene.
The more I learn about how GPUs work, the more I realise just how little I truly understand them, they're truly incredible pieces of engineering and I could watch this kind of content all day cause this was absolutely fascinating to watch.
This was a refreshing video! Thank you for uploading!
This is why Gamers Nexus is the CPU/GPU focused best tech channel on youtube. Though I've a soft spot for HardwareUnboxed too.
I just had an idea staring at the radeon vega logo on Mike's shirt. The vega emblem isn't just the two triangles. Those two triangles represent the rendered geometry of a cube. The remainder of the cube has been culled.
Thanks for making videos like this, I'll be studying computer graphics at uni this year so it's good to get a high level overview of what I'm getting myself into :D
When I play video games, I usually delete the polygons I'm not using and then stream download the ones I need. Of course, you sometimes run into problems with rgb compression textures and screen space usb ports. I usually get around this by overclocking my resolution.
You guys got it all wrong you need to undervolt your textures to improve efficiency.
yes but dont forget to overamperage the triangles before doing so.
And don't forget to filter for sinusoidal depleneration
I do some formoluted pretzelo-cakeification and ingest it in the pipeline.
Also if you have problems over clocking your resolution, try undervolting tessellation, and amp up antialiasing, it should help your monitor thermals too
If I gave the interview, I'd just smile and nod for the whole interview and hope he never asked me for my input. Then I'd ask how I even got to do the interview.
What's crazy is that I love this type of content. It means I have a goal to reach regarding increasing my knowledge of this thing I really love.
i would love so much to work with stuff like this. low level chip design is so interesting, i dont know a huge amount of it but i find the subject facinating
I geeked out over this. Awesome content as always.
very cool Steave, I only understood like 50%-60% of all that but still very cool, In fact i thought that was how rendering complex scenes worked in general, had no idea this discarding thingie was an innovation
The first thing I thought about was reflections. It could reduce this feature a lot. But still sounds very good to see optimization like this without individual games needing to do it.
I think some people are under the impression that this is a new technology, it isn't. Culling is very popular what AMD is saying is that they're improving their efficiency by doing it earlier in the pipeline if I understood correctly.
Yes, a lot of this has been around for ages on all GPUs. Some stuff is specific to AMD's hardware, but not all of it is.
OPTIMIZATION HELL YEAH!!!!
would love to see this translated into performance figures on vega. cant wait for the benchmarks
Sounds great, hopefully this tech will be more wide-spread.
That's some cool technology! Kind of makes me want to buy Vega after reviews are out
damn, the rabbit hole goes pretty deep on this one
i think i may have to watch this one a 2nd time :)
Thermaltakes ring piece glows bright..
Im watching this video stoned and all i keep thinking is. If a tree falls in the woods.... lol
Maybe the focus on GPU-side primitive discard is to fight the invisible-polygons-that-are-crazily-tesselated shenanigans that sometimes accompanies Nvidia GameWorks titles.
The big question is if this new primitive shader will be added to the standard of graphics APIs and if NVIDIA will also support it. I don't see many developers writing different rending paths for AMD and NVIDIA. It's difficult enough to work around the quirks of the different drivers to make you're engine work on all platforms. Maybe some of the big developers will do it but it would suck if it would only be used by certain games.
That's real value quality content.
AMD FOR THE WIN
"We are going to be talking about one topic today"
"How crazy my hair is during the filming of this video"!
okay so what happens if say an object that would normally cast a shadow is out of view and culled, would its shadow also disappear. So for example your in a room with a window. light is being cast from a point such as a street lamp. There is a tree between the light and the window so from in the room facing away from the window, would either still see the light from the street lamp, see the shadow the tree cast through the window, or see any light at all because the window, the tree and the light have been culled.???
This mainly depends on the game engines implementation of lighting.
Shadows and light are often baked for static object. In case of runtime computation there is not problem either because the shadow map is generated from the light-source point of view and the rendering goes through the same process
a shadow mapped light has it's own frustum when rendering the depth map. this may or may not cull object's in it.
please, more of that kind of info and GPU features.
So this is a clever extension on back face culling then. nice.
People will act like they understand what he's saying.
No idea what he said... went into a coma when he opened his mouth.
I do, most of it anyway.
what? this was pretty easy.
if you want to have your brain blown away go watch a buildzoid video or PBS space time.
i have no idea wtf he's saying..im falling asleep :D
I was thinking the same thing. Brighter minds out their than mine. hahaha
Best Tech Channel on the RUclips /End Off :D
I think you should do interviews sitten. It would be more convivial. Maybe with a beer, tea or even with a small joint if you are in a state where it's legal and your interviewee is open to it.
Fascinating listening, though missed half what he said =P
Lets hope this tech gets well used in games.
Gamer Nexus -- Introduce the Guy saying He will discuss some low level tech talk,Vega engineer -- Lets start with some High level tech talk .Gamer Nexus Damn
I'm going full red team on my next build.
Offloading the load. Genius.
Interesting stuff really, I wonder if this will boost the minimum frame rates as even some of the higher end gpus have those drops it would be great to see this optimising those scenes.
object bounding box culling is a engine thing. scissor test (for out of 2d 'frustum'/plane) and backface culling are known stuff too.. you can shovel backface culling in the dumpster if you wanna use two sided ligthing shaders (foliage. and hair). early-z culling only works efficient if the drawcalls or even just the polygons of the objects in question are rendered front to back. out of order triangles usually don't early-z either. they are hard to predict. all known things to developers tho. even me. :)
Great video. Thank you :)
i understood some of this
This is awesome!
I just zoned out and thought "damn, that's mighty fine couch!"
Old N64 games used to use the 1 sided rendering poly in engine to gain perf.
Wow mind blown.
I feel like I have a pretty good understanding about what he was saying in this interview. Great stuff but dang man. That guy was tripping over his words so hard it felt almost impossible to fallow along with what he was saying! I watched this video 3 times in order to finally get what he was saying. He spoke fairly clearly but the way he talked... It was like his voice box was buffering the whole time lol. That was a really cool interview though, it definitely continues to relate to what I keep hearing about. How things aren't getting faster anymore, just more efficient. Which sounds misleading but efficiency in my opinion is more significant then clock speeds. "A tall man can walk faster then a short man even though they both step at the same speed".
Can y'all try to clarify what the Draw Stream Binning Rasterizer and the Tile Based Rederer's relationships are to this primitive shader tech? Sounds like the primitive shader stuff is largely developer/engine reliant whereas DSBR/TBR is automatically part of the pipeline. I'd really like to know more about these things as they appear to be a big part of Vega's performance. It would appear from PC pers. testing that TBR isn't enabled in drivers yet. Also, can commands from non supported game engines be intercepted and converted to primitive shaders on the fly? If possible that could be a big win for Vega on older games.
Yes, isn't TBR supposed to be one of the things that Nvidia's been doing for quite some time? Will be interesting to see how much of a real boost that provides AMD, especially if it isn't enabled in the current drivers but will be enabled with the release drivers.
Interresting video and thats why i liked it, but considering 'gamer' is in the channel i think some basic simple info is missing to make it a good one.
What is so confusing is at what stage of processing this takes place. I believe this is a simplistic overview of the stages:
1. It starts with the CPU setting up the scene deciding based on the game where all the objects are placed and what actions objects are taking.
2. I think physics is still done by CPU
3. The GPU than wants to make an image of what the cpu sets up and loads up the enviroment into memory
4. GPU starts to render doing shadows, lighting reflections and than finally the pixels (forgot the name)
5. Sends the render result to monitor.
Now what im confused about is when primitive shader takes place. I assume during 4. GPU start to render a frame. But does it do this after having done shadows, lightning and reflections. If after how does that work if theirs a mirror behind an object?
I assume that a lot of information the GPU calculated for previous frames it reuses in the next one and adds to it like shadows and lighting.
This seems like technology that should've been included a long time ago, any ideas why it's coming up now? If the minimum reduction in rendering is 50% then surely some GPU developer would've jumped on it much earlier.
It costs transistors
John Honkerton It was first introduced in Polaris.
It used to be done by game engines.
Most game engines already do some form of culling. Unreal has a few different types to chose from.
As stated in the video, it has to be implemented in the hardware to get that kind of performance gain. Remember that they stated that if you do not implement primitive discarding before it hits the pipeline then it eats clock cycles, which means less performance. It is impossible to implement a software solution from a game dev that can access the video card at such a low level.
Steve looks like he just met (a, one of, I don't care) god(s). He seems so happy, that's cute.
I would say I understood 80% of it, but it's just because I wrote shaders and raycasting engines in the past.
When can I pre-order?
Same question
spoonikle Aug 14
pSyk Isn't that the launch day?
Don't do it! Noooooooooo
Musky Elon - No matter what, Vega will be a good buy. If I don't want it or cant afford it- ill just flip it on amazon/ebay real quick for a "refund" or "profit"
Its not like we are buying a AAA shit show of a game.
This sounds like what PowerVR was doing with Kyro. Does Vega also have tile-based rendering?
PCX and Neon 250 as well. Sega Dremcast and its arcade board variant Naomi series are currently the only tile-based rendering system hardware.
I thought that the whole idea died with Kyro. Checked wikipedia - nvidia maxwell and pascal series as well as pretty much every mobile gpu uses this technique to save memory and bandwidth.
"Does Vega also have tile-based rendering?" They are suppose to in the future. Supposedly Tile Base Render will be incorporated into software updates as the card develops. Last I heard
Not surprising from nVidia because they use the partial tile rendering method obtained from the defunct 3Dfx IP who brought Gigapixel. Note that I purposely excluded the mobile market
A question that comes to my mind when hearing about this object culling is when there's an object you can't see directly but you would be able to see it's shadow or light that it emits. is the software "smart" enough to detect that or will the shadow also disapear. For example if a building is behind your back and is casting a shadow over you would that be discarded too?
Important question: Does this need software implementation in games/engines or is it driver/hardware side?
As i have understand it it is hardware implementation but this will need software too.. Now dose this need for instance Unreal and Unity implementation that we don't know jet
i've always thought that testing whether a triangle is visiable or not, would take more time than just drawing all triangles.
Only when you are doing it through software and not hardware.
The Vega logo reminds me of Kaminas sunglasses from Gurren Lagann...
Didn't see a video on it, wondering if you got Threadripper samples? Smaller channels like Tech Deals didn't so I'm curious
Mike Mantor ... I think it would be a great name for black metal guitarist.
Listening to Mr Mantor talk was a weird middle ground of interesting and putting me to sleep
Grrrrrreat!
Doesnt Polaris have this Primitive Discard Accelerator too?
Im assuming that Vega has more changes from Fury than a re-layout for higher frequency and a more efficient shader.
DarkHalmut along with the node shrink~ tbh it sounds like a normal iteration cycle, expecting more is naive.
Navi was always more interesting on the roadmap due to the "scalable" moniker which AMD uses heavily in a context of linear scaling over multiple lithography dies. which makes it exciting. That said infinity fabric style architecture in GPU is all but inevitable and game changing, However not unique to AMD Nvidia have already announced they are doing the same. So no guarantee AMD will gain the upper hand with Navi either, especially if Nvidia beats them to the punch with their own scalable architecture.
So no early depth test discard? What a shame coz those 3 cases he mentioned are (for most engines)just a little memory bandwidth savings during vertex data fetch.
But the calculation of wether or not the geometry is in view also is costing performance (on the cpu side).
The question is if it is worth it?
Theoretically, it should be worth it in above 1440p resolutions with cranked up detail, since you rarely get cpu bottlenecks there.
On 1080p though, its probably not that worth it, since you already are more cpu bottlenecked than gpu bottlenecked.
This was incredibly boring but equally fascinating. I couldn't help but fall asleep but I also couldn't help but dream about it.
*This interview is sponsored by:* Triangles.
what if there is a reflective surface behind you that would reflect light onto the direction you are looking, would that not get culled out by this process resulting in a slightly less accurate representation of the environment?
wow
Will the driver do all of the "culling" work or is it up to game devs to code that into the game engine?
How does or does primitive discarding affect Vram usage? Does/would it help reduce Vram usage/overhead?
Edit: commented before watching the whole vid
I understood basically what he was talking about with overall goal of that tech... but.. I was mostly glazing over
Morty,.. did you get any of that :D
Rick and morty episode tomorrow
What about lighting? Does light still bounce off these objects that are not rendered? Like the back of the coffee cup for example
Lighting calculations would be done from a different perspective so the scene would be rendered again from a light's point of view depending on what light or shade is going to be visible from the player. For instance the objects that cast shadows but are not seen from the camera, you'd have to render them at some point if their shadows appear in the camera.
I've never understood the title "corporate fellow"
like "Yeah I work at AMD in the corporate division. What do I do? I don't know, but I'm definitely there. Like I have a badge and stuff. I don't have a desk, or any responsibility, but I've got a badge"
What's the difference between Primitive Discarding and Occlusion Culling ?
Shouldn't occlusion culling already be a thing? I know many games uses Umbra to hide invisible geometry.
Does this feature work native on all (past) game engines or does it have to be addressed by the game devs manually?
This also pretty much explains why players sometimes glitch outside of the map but still can see inwards, and also have their bullets shoot through it and hit things. Hit detection is only checked from one side of the object since there is no geometry rendered from the other side (that you're not supposed to see).
6:35 PervertX data? /s
I've heard that the draw stream binning rasterizer is not gonna show up in RX Vega. Is that true?
i hope they made enough RX Vegas for miner and Gamers!
COMPLETELY OFF TOPIC but this just ocurred to me...
there are researchers trying to find evidences that could point to the universe being a simulation, and this culling of objects reminded me of how quantum probability waves work.
is that map they are rendering Alborz Mountains from bf4?
am i wrong in thinking this could create lots of popup gfx if they dont get it just right?
The question is, can I disable it system wide for trouble shooting reasons?
holy crap dat old UE3 map xD
wow vega can render polygons at 350w. impressive!
this is how unreal engine 5 nanite work in ps5 Demo
yep
this will enable 8k photorealism.
Ah yes, Primitive Discarding. I hope it works well with the Turbo Encabulator.
In summary: The graphics card figures out what isn't seen, and prevents the geometry of what isn't being seen from being rendered.
Yeah, I know. But thanks anyway.
Man, the new unreal engine is NUTS!
I feel like Vega is going to have near 0 dropped frames
nvidia:
we're gonna create useless triangles on flat objects to gimp our old generations of cards and especially amd!
amd:
using great technology to discard useless triangles made by nvidia and even gain more performance in the long term.
Didn't Polaris have primitive discard accelerator though?
Yes, it's been around a while. That doesn't make it less interesting to dissect, particularly since these things tend to change generationally.