They used Just Cause 2 in the video Best. Game. Ever.

If this takes off, then there will be a reason for DRAM growth again
http://www.tomshardware.com/news/dram-memory-growth-forecast,13167.html

So they're back again. I have to say, the quality was quite impressive, but I'm still unsure if they will be able to do dynamic animation, as everything is on a grid of voxels, this makes stuff like ragdoll physics quite hard. Atomontage also looks promising, they have solved this problem by using voxels for the stationary environment, and polygons for moving objects. They even have voxel based physics implemented. I wonder how hard the lighting will be with voxels. I'm kind of worried, however, that video games that use this technology will have lots of repetitive content, due to the massive storage space requirements.

It is doable but not right now with our current technology. Although cpu speeds are increasing exponentially each year, the speed of memory is progressing at a much slower rate. Knowing the memory demands of software like this the bottleneck would be in cache misses and even more likely missing the ram itself and having to load data from the hard disk which is eons slower than the cpu. I'm betting that this technology will be in use with the film industry within the next 10 years and that have the game industry in an additional 5 to 10 years. And when it comes to difficulties such as animation, techniques are available but they don't have the perfect detail seen in the rendering engine.

The model itself doesn't use polygons, it uses voxels. Let's assume that the elephant statue is contained in a 750mm (75cm) cube. The guy said there were 64 voxels per cubic millimeter, so that makes 75 * 75 * 75 * 64, or 27,000,000 voxels. If the voxel has a 32 bit color value associated with it, that is 102MB right there. On a disk, that number can be greatly reduced by not storing information about the voxels inside the model, which nobody sees, but in memory it pretty much has to be in a 3D array so the renderer can quickly access the voxel at the arbitrary point x,y,z. (Unless the breakthrough he is talking about is in searching for voxels in another way) Actually, now that I think about it, an octree could be used so empty/unseen space need not be stored. But there is still massive amounts of data to be stored for a whole level, so I have a feeling a lot of stuff will be repetitive.

Also, did anyone notice how that guy left out games that use tesselation? So far, it has been a great solution to the polygon budget issue.

I figured that out, and now I find it quite amusing how he compared his engine to internet search engines in the first video.