Euclideon (aka. Unlimited Detail) is at it again

category: general [glöplog]

http://www.teamhellspawn.com/voxels.htm
apparently they thought voxelstein was a good idea and applied it on doom :)

added on the 2014-09-20 22:44:36 by post malone

So was that a put on accent or does he really sound like a posh toff wanker?

added on the 2014-09-21 07:35:32 by ringofyre

Quote:

The technique isn't new. However, Maxwell 2 has some new fixed-function hardware to accelerate such voxel-based techniques, and THAT is what I was referring to.
It's now fast enough for practical uses, rather than just some research papers.

Do you know any benchmarks on that? I mean a fair comparison between 780 Ti and the new cards could really help. Because you can implement that on both cards and the epic guys already showed that of more like a year ago IIRC.

Some now hardware features (DX12 level stuff I guess, e.g. conservative rasterization) are really interesting but it seems the marketing doesn't really care about those :).
Well - selling e.g. conservative rasterization to a gamer audience as great feature might be hard - agreed.

What I do not see is: more ALU power, more memory bandwidth.
Let me be a bit rough: Just from the numbers it looks like they added more power to the rasterizer units but crippled memory bus bandwidth and some other things in favor of Green IT stuff (power efficiency is nice, but I care more about the real numbers - not pixels shaded per watt). Hopefully I'm wrong here, but I have the strange feeling there must be a reason why they don't compare their new cards to GK110 cards. Looking forward to real numbers.

added on the 2014-09-21 08:29:28 by las

Quote:

Do you know any benchmarks on that? I mean a fair comparison between 780 Ti and the new cards could really help. Because you can implement that on both cards and the epic guys already showed that of more like a year ago IIRC.

Hello? Is this thing on?
There is new *hardware* in the architecture, which makes some things up to 3 times as fast as before.
Benchmarking that is rather difficult, since it will not be apples-to-apples by definition.

Quote:

Some now hardware features (DX12 level stuff I guess, e.g. conservative rasterization) are really interesting but it seems the marketing doesn't really care about those :).

Apparently you didn't get it. Conservative rasterizing can be used to accurately classify voxels, and is part of the much improved voxel performance.
Other things that boost voxel performance is that you can now have volume tiled textures, and there is multi-projection acceleration (rendering the same geometry multiple times efficiently, under different projections, such as cubemaps/voxels).

So, next time, visit ye olde clue shoppe before you speak.

added on the 2014-09-21 11:26:35 by Scali

Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight! Fight!

added on the 2014-09-21 12:48:04 by trc_wm

Quote:

There is new *hardware* in the architecture, which makes some things up to 3 times as fast as before.
Benchmarking that is rather difficult, since it will not be apples-to-apples by definition.

New hardware can be 'benchmarked' by comparing some measure (e.g. wallclock time) for an acceptably-close result. The methods of achieving the result may be different but it's still possible to make a comparison. Isn't the whole point to compare hardware? I think las is wanting to see useful measures, good descriptions of the methods, and ideally a definition of 'acceptablly-close'.

Conservative rasterization is useful indeed but in the real world it's not always the case that a new feature provides a benefit in production code.

added on the 2014-09-21 14:25:51 by fizzer

Quote:

New hardware can be 'benchmarked' by comparing some measure (e.g. wallclock time) for an acceptably-close result. The methods of achieving the result may be different but it's still possible to make a comparison.

The problem is that you are benchmarking two different implementations. So the results will be influenced a lot by the scenario you choose. As I say: it's never apples-to-apples.
And my other point was: it's obvious that the new hardware features make some things considerably more efficient, so what's the point of benchmarking anyway?

It will become interesting to benchmark once actual games start using the new features. Then you will at least be comparing a real world scenario.

Are we done now? Sheesh.

added on the 2014-09-21 14:36:51 by Scali

So you have hw help generating the dataset, but not tracing it. That part hammers the bandwidth. In my implementation experience of all this the tracing takes a lot more time than the generating and much of the generated data can be cached anyway, so id be far more concerned with side by side comparisons of the tracing - which should be valid, as theres no extra hw affecting it, right? It mainly relies on straight alu+bw, the things they are less willing to compare.

added on the 2014-09-21 16:04:21 by smash

Quote:

which should be valid, as theres no extra hw affecting it, right?

Difficult to say at this point, since they have not explained all the details of their implementation. If they also change the data format of the voxels in memory to suit the new hardware better, then the new hardware will (indirectly) affect performance here as well.
Likewise it's not entirely clear how the larger caches and prefetching algorithms affect this particular type of algorithm.

At any rate, aside from some extreme multitexturing benchmarks, the 980 seems faster than the 780Ti at pretty much all workloads at this point.

added on the 2014-09-21 16:45:51 by Scali

The internet connection of this hotel now prevented me multiple times from posting a nice post - sorry for those who would possibly have enjoyed a nice flamewar.

I just want to say: Hi smash <3

added on the 2014-09-22 00:56:10 by las

*of you

added on the 2014-09-22 00:56:36 by las

bitch please, I enjoy even mediocre flamewars

added on the 2014-09-22 10:58:59 by msqrt

Lameflar

added on the 2014-09-22 11:11:32 by Scali

Make a demo about it?

added on the 2014-09-22 14:44:16 by sol_hsa

Quote:

Make a demo about it?

A Bruce Dell-fucktro? :)

added on the 2014-09-22 16:47:05 by Scali

Quote:

Hello? Is this thing on?

No man, try it again in ALL CAPS, that might work :)

added on the 2014-09-22 19:10:09 by numtek

scali: btw, check the cost of the lookup for volume tiled resources in comparison to e.g. sampling a lowres page index map texture and looking up the highres pages in another map (the scheme i use in my raytracing structures).

the major thing i'd like to see reworked for the next directx etc is more control over compute workload balancing from within compute shaders, the ability to dispatch compute and goemetry workloads from compute shaders, etc. that would really help both raytracing performance and voxel generation.

added on the 2014-09-23 13:08:15 by smash

We're still talking about this shit?

added on the 2014-09-23 17:41:01 by gloom