I wrote in my notes: "4 concurrent, ~350 tok/s aggregate, continuous batching works." The next day I looked at completion times and saw four requests finishing one after another, 91.2 tok/s aggregate. This post shows how one division and a table of completion timestamps tells real parallelism apart from a tidy queue, and then works out how far a 98 tok/s single stream really is from the physical ceiling.
A public benchmark claims a 27B dense model hitting 2103.6 tok/s on one consumer card — about 26× faster than normal. I don't have that card and never ran that config, but three divisions are enough to show the GPU spent the run not doing forward passes. You can run all three on your own machine.
Part two of the Alkaira series. One AI agent claims to run an underground intelligence network. Another believes she controls his informants. The ledger recognizes neither. On why lying should be allowed, and how a world defends the facts.
A long-running AI society experiment: 16 characters cultivating, scheming, trading, aging, and dying inside the same demonic sect. What I saw, and why the hardest part has nothing to do with intelligence.
Four AI agents could collaborate on code, but operating the team was still a manual job. So I built one dashboard for deterministic routing, task tracking, and context-window monitoring.
Ariaki