DICE's Frostbite 2

Yeah, it's all very impressive and also gives me great hope that the console versions will look very similar to the PC.

Also interesting how fat the G-buffer is, 5 x 4 bytes altogether with the Z/stencil. It must be using tiling on the X360 but it fits in 20MB so it only has to re-send once. Gives you the impression that they've layed it out like this exactly with the EDRAM in mind...

Impressive they can do tiling along with everything else they are already running on the CPU if true.
 
Last edited by a moderator:
That's some crazy stuff in there, I have to say. Very informative, nice images, notes...

Agreed, awesome presentation. Back when I was working on the Beyond3D podcast we also looked at a presentation from these guys and they already had plans layed out for this kind of approach, and they have followed through on that way beyond my expectations. Very cool. Very interested in seeing the end results on consoles.
 
Impressive they can do tiling along with everything else they are already running on the CPU if true.
I think Laa Yosh ment that its great news that they are just below 20 megs with their frame buffer,that way they could get away with 720p because either way you HAVE to use tiling anyway so this is great news.
 
Wait...do I see FP16 HDR output under 'SPU shading overview' in that slide ?

What's wrong with 16-bit float for the SPUs ? ^_^

The highlights for SPUs are:

* SPU Translucency:
dicesssperf.jpg


* SPU Culling
screenshot20110308at124f.png


* SPU + RSX Light-Tile Deferred Shading
screenshot20110308at124.png


screenshot20110308at124.png


EDIT: Wondering if they have enough resources left for more SPU post processing. e.g., MLAA, object motion blur, DoF, etc.

Not bad for a CPU eh ?
 
Mind you tiling is not confirmed yet :) but it's the only way I can see such a big G-buffer working for them, so it's a safe bet.
And yeah it looks like they've said, "so we're gonna have to tile because of the extra RGBA radiosity buffer, then at least make full use of the doubled EDRAM size and fill out the rest of the memory with useful info". Hence the different extra channels associated with material properties and the lack of trickery with the surface normal channels.

If it really is the case then it'll also be interesting to compare it with Crysis 2 because they're both multiplatform engines developed to also utilize more powerful PC systems while not sacrificing key features on consoles; and the feature sets are also very similar, but there are some differences in their implementations.
For example Crysis fits the entire G-buffer into the X360's EDRAM and running at similar frame rates means less total resource utilization - so is it the fully dynamic GI that requires those extra shader cycles and bandwidth?

I realize that this can take us into dangerous territories and can spawn leftover threads, but there's far more logic behind this one, considering the far more similar goals and restrictions of the two engines' developments...
 
I think Laa Yosh ment that its great news that they are just below 20 megs with their frame buffer,that way they could get away with 720p because either way you HAVE to use tiling anyway so this is great news.

Oh I see, that makes sense.

Excuse the idiot question, but does this mean you would always need tiling when running a 720p deferred renderer due to the size of the G-buffer? I'm not as up to date on deferred set ups as I am on the typical forward renderer, would love to learn though.
 
I think multi-platform engines are superior to exclusive engines nowadays. I thought I would never say this, this generation.

Probably depends on what they want to achieve. The KZ3 engine only supports LDR and non-destructible (or partially destructible) environments. It also does "similar" SPU occlusion and culling, and RSX + SPU deferred rendering as FrostBite 2. It turns its attention on stereoscopic 3D, MLAA and other SPU post processing instead.

I heard animation and AI are impressive on both games ? (True ?) Also need to look at close up facials, hair, etc.

We need one more round of exercise to see cross-pollination, and may be more innovation. This is [strike]good[/strike] great !
 
EDIT: Wondering if they have enough resources left for more SPU post processing. e.g., MLAA, object motion blur, DoF, etc.

Not bad for a CPU eh ?
They 40ms split between 5 SPUs, right? That's 8ms latencies aside, then their early cull is super fast using SPU assembly, Mlaa as in GoW3 is 20ms split between 5 SPUs, even taking in account lately and as they scheduling seems top notch they may have quiet some muscles left for other tasks.

Pretty impressive when I see their use of the Cell is makes me wonder about this topic I opened a while ago or the talk that happened in the handled forum in the ipad2 thread (especially Panajev2001a statements). Hardware is a fun thing and a tough job so many options I'm believer that what the PS360 achieve could have be achieve for way cheaper hardware wise (possibly a big fusion chip) relying on more fixed function hardware as long as you design it to run the proper software model.
 
I think multi-platform engines are superior to exclusive engines nowadays. I thought I would never say this, this generation.

with the lighting, animation and destruction of C2 and BF3 I have to agree. Give edge to FB2.

They appear poised to have amazing effects realized on machines that I never thought possible on 6 year old tech.
 
They 40ms split between 5 SPUs, right? That's 8ms latencies aside, then their early cull is super fast using SPU assembly, Mlaa as in GoW3 is 20ms split between 5 SPUs, even taking in account lately and as they scheduling seems top notch they may have quiet some muscles left for other tasks.

Pretty impressive when I see their use of the Cell is makes me wonder about this topic I opened a while ago or the talk that happened in the handled forum in the ipad2 thread (especially Panajev2001a statements). Hardware is a fun thing and a tough job so many options I'm believer that what the PS360 achieve could have be achieve for way cheaper hardware wise (possibly a big fusion chip) relying on more fixed function hardware as long as you design it to run the proper software model.

Some of their slides mentioned 6 SPUs (probably occasional use). I was wondering what they are up to. I don't know what to say when everyone forget to talk about the PPU VMX. I guess it's used for mundane work ?
 
To be honest, back in the early years of the PS3 architecture discussions I've totally dismissed the idea of doing the actual shading on the SPUs, for reasons like problematic texture access and the very small local memory's limitation on tile size. It just looked like the implementation would be far to low efficiency and work intensive for it to make sense.

So it's really amazing to see how Dice has still gone all along the way to utilize their processing power; and it's kinda evident now that deferred rendering can remove the texture access problems completely.
 
Excuse the idiot question, but does this mean you would always need tiling when running a 720p deferred renderer due to the size of the G-buffer? I'm not as up to date on deferred set ups as I am on the typical forward renderer, would love to learn though.

On the X360 it is very likely, because for a fully deferred renderer you need the following at least:
depth 3 bytes, usually combined with stencil 1 byte
normals 3 bytes
diffuse albedo 3 bytes + specular whatever 1 byte

Theoretically you'd have an extra byte aside the normals but you need to work with 4-byte wide data because of the hardware's layout. So you can usually stuff some other kind of data into the remaining byte, like material identifiers, occlusion and so on.

Anyway you'll need at least 3 full sized buffers, but the 10MB EDRAM can't fit them in full 1280x720.

Trials HD cuts off 20 rows, and fills them with black; and Crysis isn't fully deferred and skips the diffuse albedo channel. Dead Space 2 is fully deferred as I hear and so they're probably using tiling.
 
Pretty impressive when I see their use of the Cell is makes me wonder about this topic I opened a while ago or the talk that happened in the handled forum in the ipad2 thread (especially Panajev2001a statements). Hardware is a fun thing and a tough job so many options I'm believer that what the PS360 achieve could have be achieve for way cheaper hardware wise (possibly a big fusion chip) relying on more fixed function hardware as long as you design it to run the proper software model.

http://twitter.com/#!/ChristinaCoffin/status/45163944628862977

Hints that they might describe what was done on the 360 at SIGGRAPH (my last visit was in 1989, so I'm overdue for a conference visit), which I think will be very interesting to see how they utilize different hardware to achieve the same results.
 
Last edited by a moderator:
Some of their slides mentioned 6 SPUs (probably occasional use). I was wondering what they are up to. I don't know what to say when everyone forget to talk about the PPU VMX. I guess it's used for mundane work ?

Does the PPU on the Cell have any VMX units? I thought that was only on the Xenon.

On the X360 it is very likely, because for a fully deferred renderer you need the following at least:
depth 3 bytes, usually combined with stencil 1 byte
normals 3 bytes
diffuse albedo 3 bytes + specular whatever 1 byte

Theoretically you'd have an extra byte aside the normals but you need to work with 4-byte wide data because of the hardware's layout. So you can usually stuff some other kind of data into the remaining byte, like material identifiers, occlusion and so on.

Anyway you'll need at least 3 full sized buffers, but the 10MB EDRAM can't fit them in full 1280x720.

Trials HD cuts off 20 rows, and fills them with black; and Crysis isn't fully deferred and skips the diffuse albedo channel. Dead Space 2 is fully deferred as I hear and so they're probably using tiling.

Much appreciated my good man.

Wonder what's going on there that it needs to wait until Siggraph. :(

Maybe they only had enough time to talk about one console? Makes sense to tackle what's being done on the ps3 since that seems to be the harder nut to crack for studios.
 
Back
Top