Spectral-AI uses RT cores for MoE routing

This is a clever hardware-co-design experiment, but the headline number is routing-only, so the real value depends on how much of total MoE latency the gate actually consumes. The more interesting claim is the specialization finding: if experts cluster by syntax rather than topic, a lot of “semantic expert” intuition in MoE land needs revising.

• –The 218x figure is for routing, not full-model inference; the repo’s own framing is more conservative than the Reddit headline.
• –Reported router accuracy at 95.9% and a +1.5% perplexity hit suggest the approximation is usable, but not free.
• –The approach matters most for very large expert counts, where O(N) routing can become a genuine bottleneck.
• –The syntactic-specialization result is the strongest research angle here: it has implications for interpretability, routing design, and expert editing.
• –This looks more like an open research platform than a finished serving product, which is why the open data and paper trail matter.