Tech Analysis: Battlefield 3 PC

For many PC gamers there is a nasty stigma attached to multi-platform releases. The common criticism is that the development of games on the PC platform, not to mention the commercial uptake of GPU hardware tailored towards new technologies such as DirectX 11, has been hamstrung by developers' growing attention towards the console platforms over the years.

Of course, this is the natural result of a marketing drive inherent in publishers who, in a bid to maximise their potential audience, continue to encourage developers to cater their fundamental game design choices towards much older console architecture. Sadly, this is so often followed by an unambitious, token PC release trickling in a few months later, showing little in the way of technological progress over what could have been achieved years ago on the platform.

Meanwhile, with both Sony and Microsoft are still evidently committed to keeping their current generation of consoles in contention for as long as possible, it would seem that using the PC to showcase the possibilities of next generation hardware is incongruent with the notion of creating a successful multi-platform title. Fortunately, Battlefield 3 is now here to make that stand, and to prove that it's possible to have it both ways.

As shown by our recent Battlefield 3 Face-Off feature, series developer DICE has achieved something special here on all three platforms, proving that they've honed their multi-platform juggling skills further since their last release. Previously on the PC side, criticisms were levelled at Battlefield Bad Company 2 for its smaller environments and a limit of only 32 players maximum per server, which rang false to fans who were used to the 64 player capacity of previous games. Promises were made at the start of Battlefield 3's development, however, that the PC version's 64 player count would make a comeback, and that it would be the one to take the greatest advantage of their newly revised Frostbite 2 engine

The use of DirectX 11 provides the basis for some of Battlefield 3's stand-out graphical features - the API's Compute Shader technology powers the wonderful lighting.

The Rise of DirectX 11

Just last month, DICE's rendering architect Johan Andersson explained the finer details of this upgrade during a presentation at the GeForce LAN 6 conference. The first sign that things were working in the progressive PC gamer's favour was the announcement that support for Windows XP and DirectX 9 was in fact being dropped entirely.

A clean break to DX11 offers many benefits in performance, including optimised streaming, Compute Shader support, and Order Independent Transparencies (OIT). On the first point, Andersson describes how in older versions of DirectX "there would be CPU spikes as it streams in new content, because the APIs weren't designed for this approach. We've been working pretty closely with Microsoft and NVIDIA to come up with a solution in DX11 where you can have a separate thread that creates all the textures and meshes, and the drivers are specifically optimised for this."

The net result of this, he continues, is that it "makes sure that these giant textures are loading or uploaded in parallel so you don't see any CPU stalls."

Such improvements to the rendering pipeline inevitably entitles DICE to more creative breathing room when it comes to designing densely-packed levels. Multiplayer stages are now strewn with objects that can be instanced; a process whereby repeated rendering of the same object across a level, from plastic baskets to identical buildings or warehouses, can be achieved within a single draw call to the GPU. This means that the viewing distance on objects can subsequently be increased for all graphical configurations.

A significant benefit to utilising DX10 and 11 exclusively is the way it enforces a consistent visual standard across all graphical settings. Although the single player campaign offers a much more restrictive, linear experience compared to the multiplayer, the above video shows how closely matched these settings can be in like-for-like situations. Our mid-level PC build, equipped with an Intel Q6600 processor and a DX11 compatible Radeon HD 6870 card, could achieve stable frame rates on anything up to the "Ultra" graphical bracket without issue.

Perhaps more impressive though is how all geometry meshes and character models look by comparison even on the lowest setting, with only the dithered shadows being an obvious point of distinction between the settings. Minor objects on the floor may take a massive nosedive in texture resolution, much like the floor texturing itself, but they all stand present and correct with no omissions. It's the same case for foliage on trees, and the smaller sprites used on grass tufts, which services the fairness of the multiplayer mode where any such element could be used for cover.

Comparing Ultra with the Low graphical presets shows us that all objects and foliage are in place regardless of setting - although a price is paid in texture quality. Flickering, dithered shadows also stand out as a larger distraction on the lower settings.

On the multiplayer front, the decision to go with these new APIs has a massive impact on the actual competitive playability of the game. For example, another of the major additions offered here are Order Independent Transparencies, allowing for effective blending of multiple layered transparency effects, particularly in cases where they overlap and interact with each other. Once written out to memory, these layered effects can also be blended with atomic particles, such as ash or sparks, giving a much busier look to indoor scenes. Much like the many post-processing effects used by the game, this obscures the player's view in a realistic manner. Also interesting is that smoke has real volume in Battlefield 3, to the point where it casts shadows.

In practise, this means that looking through a sequence of two or more muddied glass panes with lit particle effects flying about between them resolves itself without errors. An attempt to sidestep the issue for the sake of DX9 support could have invoked crude simplifications or even omissions to the visual make-up of the game, which wouldn't have been ideal for an even multiplayer experience. Much like the notion of removing any foliage or objects in a scene, such a clinical, pared down look would have given those running on lower settings an unfair advantage.

Much like in Bad Company 2, tessellation is also a staple feature in the DX11 incarnation of the game, though the implementation is more accomplished in the new game. This provides extra 3D definition to textures by approximating a small mesh around portions of otherwise flat surfaces, or elements of geometry. The result here is well worthwhile. Cobblestones and ruined roads now show convincing peaks and troughs in formation, while meshes on characters are much more distinct.

While the campaign experience - on the early levels at least - is similar between console and PC, later levels and many of the multiplayer maps utilise larger, expansive environments which offer up a significant level of improvement on PC.

Lights, Camera, Action

These optimisations also extend to lighting, which is often cited by DICE as the biggest step forward for the engine. It is now capable of many more light sources thanks to Compute Shader support, which presents a new way for programmers to write for graphics hardware. The idea is reminiscent of Killzone 3's deferred shading, but in this case the GPU is being treated like a parallel CPU, with the lighting implemented in a separate pass to the main rendering.

The rendering pipeline now sees the general geometric layout of a scene being sent to the GPU as a large G-buffer, at which point the tile-based lighting model fills in by the pixel at a later stage. This is rather than addressing each object with a lighting in one concurrent pass, as per the conventional solution. Thanks to the culling of unneeded light sources afforded by this selective approach, this allows for everything from emissive particles to specular highlights to run in tandem whilst saving greatly on RAM bandwidth on more restrictive cards.

Emissive particles function as floating light sources, which greatly impacts the visual make-up of the environment, creating a much more realistic effect.

The importance of such a streamlined method of rendering lighting to a scene is not to be understated, and even manifests itself in particle effects such as ash or dust. Not only can these particles reflect and inflict shadows of their own accord as they float mid-air, but they can also serve as small dynamic, light sources themselves.

DICE's Johan Andersson has much to say on the matter, explaining that "we can also have destructible light sources, and animating light sources. The cost of the lighting in a scene is more dependent on the number of pixels the lighting covers, rather than the number of objects it covers. So we can have a highly flexible lighting scheme here."

This new approach realises a game experience that wasn't possible before, with even flashlights and laser-sights now producing a blinding bloom effect over the player's field of view. For DICE, the principle of bloom is very much considered a part of the High Dynamic Range (HDR) pipeline, whereby if a light source exceeds the average contrast value in any given environment, it bleeds over. Although its implementation on flash-lights can be slightly overbearing at times, particularly during levels set during daylight where they perhaps shouldn't have so great an influence by contrast to the global illumination, it is still impressive to have the effect implemented into the run of play.

Advanced Anti-Aliasing Support

There are subtle drawbacks to using HDR post processing such as this, however, as it is factored in at a much later stage in the processing cycle. In particular, this has an impact on the effectiveness of conventional Multi-Sample Anti Aliasing (MSAA), which is oblivious to how the final image is effected after its application. This is where DICE's inclusion of Fast Approximate Anti Aliasing (FXAA) option in the graphics settings becomes relevant. As another form of post-processing that functions at a pixel level, it can soften the high contrast edges left on decorative elements such as fences or foliage.

Offered at three levels of intensity, including low, medium and high, FXAA softens the overall image ever so slightly, but has a keen eye for the jagged edges left on 2D elements especially - an area that sample-based AA techniques can't detect easily. It also does a good job of softening the straight edges without being too much of a drain on the GPU's resources. Given that 4x MSAA incurs a massive increase to the required fill-rate on a graphics card, this alternative solution does a very respectable job of cleaning up the image on a much lower processing budget.

No anti-aliasing in the first shot and then we apply the various forms of AA the PC version of Battlefield 3 offers. The opportunity to utilise both post-process AA and 4x MSAA in the Ultra mode provides the best of both worlds.

That being said, by offering both options to the player there's the possibility of reaping the benefits of each by having them run in tandem. While FXAA's pixel-level approach may pick up on jaggies left by leaves, power lines and crate grills, MSAA's ability to analyse edges at a geometry level means that it can catch out smaller, thinner details in objects that wouldn't be detected by the former method. So by having both running in conjunction at the highest settings, a more thorough dissection of each frame's high contrast edges is possible, making for a remarkably smooth image.

Physical Effects

Another major point of revision for Frostbite 2 is its proprietary physics and destruction engine. Above all else, the changes made here have the most tangible effect on how a multiplayer game plays out, with chunks of a concrete wall being procedurally chipped away to reveal sniper positions, or the flooring giving way entirely to a grenade blast. This places you in an environment that is in a constant state of geometric flux, making the level you finish playing on quite different to the one you started.

Much like in Bad Company 2, the precision of this chip effect is based on small pre-defined chunks in the geometry. Fortunately, instancing of objects also applies to these composite meshes - buildings with multiple destructible parts - which means that the reduced state of each object is stored separately for each instance, and that the performance benefits are maintained.

The campaign mode offers a far less ambitious approach though, featuring a mostly pared down implementation of the destruction engine which can sometimes make it difficult to distinguish between walls that can crumble to gunfire and walls that will remain eternally upright. Similarly, the physics engine effectiveness on small objects can be inconsistent by comparison to the multiplayer's all-encompassing approach, with grenades animating minor decorative parts like plates and the surrounding chairs, while tables and stacked crates remain inexplicably static.

It's a shame that this mode couldn't follow in the footsteps of Bad Company series' more free-form single-player structure, which involved the player in larger sandbox-style environments during certain battle sequences. As it is, the multiplayer mode demonstrates the Frostbite 2 engine's technical virtues far more admirably when it comes to physics.

DICE's destruction model is utilised extensively for the showpiece engine-driven cut-scenes in the campaign, but it's in the multiplayer mode where the game really shows off how versatile and powerful this technology is.

Beyond the Call of Duty?

Although the inclusion of a single-player component may come across as a rudimentary, tit-for-tat rebuttal to Call of Duty's full-featured offering, the fact of the matter is that DICE has used the campaign as an opportunity to showcase their new technology effectively. Save for the restrictive destruction physics, it directs the player's attention to some spectacular visual improvements, be it their new lighting engine which allows for an increased number of light sources, or the subtler performance optimisations resulting from Battlefield 3's native DX10 and DX11 support.

While some may argue that multiplayer maps such as Operation Metro still hint towards the influence of the console versions' 24 player count, the bulk of the experience is quintessentially closer to the sandbox map design of its progenitor: Battlefield 2. With that said, the plight of the PC developer shouldn't just be predicated on recreating the platform's halcyon days, but about discovering new ways to promote technology which to this day remains largely untapped. With Battlefield 3, it's satisfying to for once find a multi-platform game that still sets the PC version aside to show what could be possible in the next generation of hardware.