Biggest advantage is response time on complex views, followed by the ability to guarantee pixel-perfect presentation on any device capable of websockets and jpeg decoding. Caching of content is on a final tile basis rather than full viewport. This increases the hit-rate substantially, as tiles are not at fixed pixel boundaries, but rather aligned with logical UI elements. Fixed sized elements will more likely produce the same final tile(s) with equivalent SHA256 regardless of client device. Hashing is performed in terms of the UI element's properties (dimensions, size, text, color, etc), so cache lookups can be made along these lines.
Biggest downside is latency - This approach has a similar practicality envelope as Google Stadia, et. al. WRT distance from datacenter. 2nd biggest downside is the fact that this technique is fundamentally incompatible with classic web development ideologies and requires new ways of thinking about solving problems. I personally find this last point to be a benefit, but I recognize it would prevent wider adoption in the community.