Width and depth are the two canonical transformer scaling knobs; this paper proposes branching — parallel composition of full transformer blocks — as a third, backed by scaling laws and H200 hardware benchmarks. The 51% throughput gain and 34% KV-cache cut are concrete, reproducible claims that will drive immediate follow-on work across labs.
Original headline:Branching: A Third Transformer Scaling Axis With 51% Throughput Gain
Track only the AI that matters to you
Your own agent, watching your companies and topics.
Build your agent →
We use essential cookies to keep the site working (login, form security). With your permission, we also use analytics cookies to understand how you use the site.
Privacy policy