SIMARC runs on privately owned bare metal, operated directly rather than rented: exhaustive combinatorial search, SAT and algebraic decision procedures, quantum trajectory simulation, large-cohort statistics. Sustained equivalent capacity costs one to two orders of magnitude more on demand, and that gap is the entire point — the research style depends on running exhaustive censuses without per-hour deliberation.
The cluster is deliberately asymmetric. Some capacity finishes a single hard instance in reasonable wall-clock time; the rest is throughput hardware that is cheaper, wider and substantially slower per core. Work is routed to whichever shape the problem actually has, and shard counts are weighted by measured throughput rather than split evenly — an even split leaves the fastest machine idle for most of a run.
Memory, not cores, is usually the binding limit, and it is a scheduling input rather than an afterthought. It has driven architecture decisions rather than optimisations: where a dense representation exceeds what any single node can hold, the program is rebuilt around one that does not.
The cluster runs a small purpose-built scheduling layer rather than a general workload manager, for one reason: the expensive failures here are the invisible ones. A job that stalls, crash-loops, or runs to completion having silently done nothing costs more than a job that crashes, because nobody looks. The layer exists to make those states loud.
No long job begins without a bounded preflight that measures throughput and refuses to bless work whose extrapolated wall clock exceeds its declared budget. Every long job is detached, survives logout and reboot, and checkpoints to persistent storage. Ambiguous states are escalated rather than killed — a false positive that destroys a twelve-hour proof is worse than a stall.
Fleet inventory, addressing and the scheduling implementation are operational detail and are not published.
The scheduling layer is general — it dispatches any embarrassingly parallel decision procedure — and the cluster has spare capacity. SIMARC will host external exhaustive-search problems that fit the doctrine: bounded, checkpointable, preflight-able, and prepared to ship a certificate. Formal intake opens in Phase 3; until then the conversation starts by email.
Scaling is a purchase decision rather than an architecture change: the scheduler already dispatches by measured throughput, so new capacity joins by shape.
