Can 1,000 different computers become a supercomputer?

Which jobs can travel across mixed machines?

Begin by distinguishing independent tasks from tightly coupled computation. Rendering separate frames or processing separate documents can tolerate uneven nodes; a simulation exchanging boundary values at every step may wait for its slowest participant.

Inventory instruction sets, RAM, accelerators, operating systems, drivers and network paths. Partition compatible resources, stage datasets, and schedule jobs against explicit requirements. Slurm can coordinate heterogeneous job components; it does not combine physical memories into one transparent address space.

Why a thousand workers do not mean a thousandfold gain

Amdahl’s idealized model is S(N)=1/((1−p)+p/N). If 99% of a fixed task is perfectly parallel, 1,000 equal workers yield about 91× speedup, not 1,000×. At 99.9%, the bound is about 500×. These are calculated upper bounds before communication, imbalance and failures, not measurements. For mixed machines, measure completed work per unit time and scheduling overhead rather than treating every node as an identical worker.

Grow from a measured pilot

Pilot representative workloads on 8, 32 and 128 nodes before expanding. Record throughput, tail completion time, energy per completed job and recovery behavior. One thousand assumed 250 W nodes already imply 250 kW of IT power before networking and cooling.

Software compatibility is a separate gate: Open MPI 5 documents limitations for mixed data representations. Check the exact stack, isolate incompatible architectures, use checkpointing and reserve spare capacity. An economical grid for independent jobs may be a better outcome than a tightly coupled supercomputer.