For those more technically inclined
http://graphics.cs.williams.edu/papers/CloudLight13/Crassin13Cloud.pdf
No wonder Epic ditched Voxel cone tracing in UE4, it is a real performance hog.
"We consider the town or city-wide scale to be the most plausible
deployment scenarios. One could envision existing Internet access
and content providers extending their local server nodes with 3D
rendering capabilities in the same way that they already have ex-
tended those nodes to support media caching, digital goods stores,
video streaming services, and massive multiplayer games. Our
scalability results are reported for the college campus network be-
cause it provided a network under real-world load circumstances on
which we could physically manage 50 client machines."
"The choice of global illumination algorithm has major implications
for target bandwidth and client devices. Our voxel approach allows
extreme amortization of resources and places almost no computa-
tional burden on client devices. However, voxels do not scale to
large scenes. We mitigate this limitation via a multiresolution ap-
proach, underestimating distant indirect illumination in the process."
"Irradiance mapping requires the lowest bandwidth of our algo-
rithms, with latency lower than voxels due to utilization of client
computational resources. It also integrates easily into existing game
engines. Unfortunately, irradiance maps require a global geometric
parameterization. While decades of research have provided a mul-
titude of parameterization techniques, these do not address prob-
lems specific to global illumination: handling light leaking where
texels lie below walls or keeping world-space samples close in tex-
ture space for efficient clustering into basis functions. We see the
authoring burden of parameterization as one reason developers are
moving towards other techniques, e.g., light probes."
"In comparison to irradiance maps, photons require significantly
more bandwidth and client computation; however, they eliminate
the need for a global parameterization and allow smaller, progres-
sive updates that enable fast reaction to significant scene changes."
"For our scenes, we tuned batch size to maximize transferred lighting
without delaying updates more than a frame over base network la-
tency. With increasing client connections, we reach maximal band-
with quite quickly; while we can easily afford emitting additional
photons on the server, this would increase bandwidth (and hence
latency) for every batch of photons sent to clients. This suggests
we could afford additional computation to select more intelligent
photons, e.g., using photon relaxation to compute better photons.
Alternatively, finding an optimal photon compression would reduce
per-photon bandwidth. This is a somewhat unique situation for in-
teractive global illumination techniques: having excess computa-
tion to spend on optimizing sampling"
@TheVoxelman on twitter
