You’re hitting the sweet spot between speed and cinema‑grade, but it’s a tightrope. First off, lock down a single light source—point or directional—so you can use a physically‑based BRDF that won’t explode on every frame. Next, ditch full‑resolution GI; use a voxel cone tracing or light‑map baked into a few spherical harmonics. That gives you the global ambience with only a few samples per pixel. Then, keep your shading model lean: a single Cook‑Torrance with a prefiltered specular map, no per‑pixel fresnel hacks. Use screen‑space reflections only for the highest‑importance surfaces; everything else can use a cheap planar or cube‑map fallback. Finally, keep your shadow maps at a modest resolution and drop cascade count if you’re hitting 60fps. If you can pull that together, you’ll still look like a movie while staying in real‑time.

Glad you’re on the same page—remember, every half‑precision tweak is a gamble. If you can get those normals in 16‑bit without clamping, you’ll save a lot. Also, keep the light’s intensity in a small range, then bump the specular fresnel with a single texel lookup. Test in a packed scene; if the shadow maps start to bleed, swap to a cascade‑less approach and use a temporal filter. Keep the pipeline lean, keep the eye happy, and don’t let a single stray sample ruin the frame. You’ll be smashing 60fps while still looking like a blockbuster.