What is common knowledge in your field, but shocks outsiders? GPUs can do insane triangle counts these days. What they can't do is insane numbers of objects. That number hasn't grown much since the noughties. A million triangles per object? No problem. A million objects? You ask the impossible! - ThreadSky

tomforsyth.bsky.social • 4 days ago

What is common knowledge in your field, but shocks outsiders?

GPUs can do insane triangle counts these days. What they can't do is insane numbers of objects. That number hasn't grown much since the noughties. A million triangles per object? No problem. A million objects? You ask the impossible!

Reposted from Alan Wolfe

What is common knowledge in your field, but shocks outsiders?

(As a researcher) Getting to the edge of human knowledge is not a very far walk.
Once you are there, you dont have to be a genius to add to it.

Comments

pedrothedagger.bsky.social•4 days ago

Now, an insane number of instances of the same object, that’s a different story!

ravikanodia.bsky.social•3 days ago

I love the revelation that it turns out humans are essentially incapable of noticing the similarities between a hundred otherwise identical trees, if they are rotated to face different directions and perhaps scaled slightly up or down

tomforsyth.bsky.social•4 days ago

Ah, that cursed trickster of many faces, but never the one you want - the "instance".

cpruett.bsky.social•4 days ago

jeroendstout.bsky.social•4 days ago

This sadly shocks a lot of insiders, too 😬

demofox.bsky.social•4 days ago

Do you think we will ever get an api call to rasterize a tlas?
If so, would that help?

tomforsyth.bsky.social•4 days ago

No, and no.

tomforsyth.bsky.social•4 days ago

To elucidate - graphics thingies that don't have a clear rigid spec across multiple vendors have typically died a horrible, agonizing and completely deserved death.

And there's better approximations that the app can use that are tailored to the specific situation they find themselves in.

hikey.org•3 days ago

@jbikker.bsky.social - Thought experiment: How would this look like in tinybvh? Maybe passing a frustum and a callback?
But how could it group triangles? Maybe a callback per biggest blas box that fits in the view and their transform.
But sorting by material would still be an issue.

hikey.org•3 days ago

Naively, one callback per triangle may work - but would it be useful beyond debugging?

demofox.bsky.social•3 days ago

My thought was: if part of the problem is all the draw call commands, a tlas has all the objects to draw already, instanced and oriented. Saying "draw this tlas" is just one command and the rasterization could be the same as usual.
Using a tlas to improve rasterization itself sounds interesting too!

hikey.org•3 days ago

"the same as usual" is doing some heavy lifting here (:
You'd still get a soup of randomly sorted primitives. That would be hard to render.
The usual way normally has things already grouped by some metric.

demofox.bsky.social•3 days ago

Im not a hardware guy, but we make blas's from VBs that are the same format as what we'd send to the rasterizer and usually still have access to the raw vertex data after doing a ray trace to get vertex data.
It seems a draw call against a tlas could be the same as a draw call against each object.

jonbrodsky.bsky.social•3 days ago

are "objects" a stand in for drawcalls here? also, is this specific to non-mobile gpus, or does this rule hold true across all modern gpus?

tomforsyth.bsky.social•3 days ago

See other sub-threads for the discussion.

kaynsd.itch.io•4 days ago

The thing I remember from like a decade back is that everyone was psyched to make millions of instanced meshes to build these gigantic worlds full of hundred of nearly identical people and places... and then every game design went on deep personalization and every engine programmer started panicking

stormancer.com•4 days ago

"And each tree is going to be slightly different!"
Me, suspicious: "what do you mean, by 'slightly different' ?"

obiwansotti.bsky.social•3 days ago

Why is that? I fel like iRacing suffers from this and would love to understand it better.

tomforsyth.bsky.social•3 days ago

https://bsky.app/profile/tomforsyth.bsky.social/post/3lrrwbfym2k2i

obiwansotti.bsky.social•3 days ago

More of a deep pipelining issue then?

That is a hard problem
To fix.
Shorten the pipeline, lose clockspeed. Cache doesn’t fix it. Just kinda stuck.

Is that right?

tomforsyth.bsky.social•3 days ago

Exactly. "State" = a form of cache. Changing state = cache invalidation/tracking. A famously fun thing to parallelize :-)

nothings.bsky.social•4 days ago

What's the barrier to a million objects?

tomforsyth.bsky.social•4 days ago

Pipeline stalls and/or flushes. Similar problem as branch misprediction. When you have a zillion things in flight and you need to change some state right in the middle of all of them, there's a ton of overhead and synchronization and pipeline draining that has to happen.

rattenhirn.bsky.social•4 days ago

Only if you are fancy and want to change the heavy states in between objects! ;)

tomforsyth.bsky.social•4 days ago

That's the definition of "object" versus "instance", yup.

quadraticink.bsky.social•4 days ago

Fortunately, instancing keeps getting more flexible. In practice, I don't really see cases where you need that many objects and they can't be instanced.

kaynsd.itch.io•3 days ago

That was the thinking a decade ago where everyone was thinking every character in a scene would be a variant of a base mesh with a different t-shirt and they could animated huge crowds and engines were made with that pipeline in mind
Then Fortnite happened and everyone wanted everything personalised

nothings.bsky.social•4 days ago

but drawing a million objects doesn't require a mllion state changes. vertex shader can pull object info.

tomforsyth.bsky.social•3 days ago

That's the various flavours of "instancing" and it usually has some tough limits.

lisyarus.bsky.social•3 days ago

Could you elaborate on these tough limits?

tomforsyth.bsky.social•3 days ago

You choose them. Same textures? Same vertex formats? Same number of bones (sometimes zero)? Same shaders? Same index/vertex list?

The more limits you impose, the faster it goes, but the more irritating they are to work within. Fewer limits = easier to work with, but each change stalls more.

archo5.bsky.social•4 days ago

I've sometimes wondered how many objects could be rendered with bindless everything + single vertex format

either if assuming that there's just 1-few shaders, or if it was possible to use many different shaders with a RT-like table of shaders

tys.work•3 days ago

Lately I have been trying to understand why a million triangles is fine unless the triangles are smaller than 1 pixel, and what the difference is

tomforsyth.bsky.social•3 days ago

Triangles less than about 10 pixels aren't any slower PER TRIANGLE. They're just not any faster, even though they're smaller. But because they're smaller, you need more of them, so overall things get slower.

tys.work•3 days ago

Oh I see! I saw some video that was like "LODs work but not how you think" that seemed to suggest that polycount is irrelevant but there was some sort of evil magic happening with wavefronts/fragments when many triangles occupied one of them. I am far from a graphics programmer and just a dabbler

crispclover.com•3 days ago

Did you mean this video? https://youtu.be/hf27qsQPRLQ?feature=shared

tomforsyth.bsky.social•3 days ago

That's a great video (although my usual gripe of using FPS rather than msPF... sigh). There's a reason that 10 pixel tris are about the sweet spot, and this has been true for around 50 years, even though the actual rendering tech has changed insanely over that time.

tys.work•3 days ago

yep thats the one

kerberjg.bsky.social•4 days ago

Silly question! Would you say that combining a lot of those objects into a single mesh (assuming they're static) could be a significant optimization?
Or is it more about combining draw calls?

tomforsyth.bsky.social•4 days ago

You can do quite a lot of draw calls, as long as you change almost no state in between them. In that sense, it's not that different to combining them into bigger meshes, just marginally more convenient.

Comments

Posting Rules

Reply