Got it, I see, a bit of background then:
lumen GI and reflections was designed to scale from low-end PCs to modern consoles to the RTX-powered supercomputers of today. To compute indirect lighting, it uses ray-tracing, but it traces rays on two different scenes: there’s the cheaper ray-tracing called software ray tracing (SWRT) and then there’s the more expensive hardware-accelerated Ray-tracing (HWRT). SWRT traces rays against signed distance fields (SDFs), which are blobby representations of the scene that are fast to trace against. Since they don’t match up with the geo 1-1, you can have light-leaking and missing detail.
To trace against Signed distance fields still means tracing against a bunch of individual objects, but you can also merge every mesh’s SDF into a low-res global distance field that’s much cheaper. Detail tracing tests individual mesh SDFs, global tracing is just the global distance field.
Hardware ray tracing meanwhile can intersect against the same geometry that the Gbuffer has (with a few caveats), so it’s generally more accurate but much more expensive. That said, there is no reason your SDF scene should be performing that terribly, especially with a small scene and a tame instance count. Do you have an RTX card?