/HARDWARE
Kova-ready hardware
What matters: CPU threads, GPU partitioning (MIG/MPS), RAM, and bandwidth. Build for utilization, not vanity specs.
/01
CPU
More threads = more schedulable slices. Favor sustained clocks and memory bandwidth over bursty marketing numbers.
/02
GPU
MIG/MPS capable GPUs enable reliable fractional scheduling. Partitions reduce contention and improve isolation.
/03
Network + storage
Throughput dies on slow disks and congested links. Prioritize NVMe and stable upstream bandwidth.
/SINGLE BOX
Starter
- 8–16 CPU cores
- 64–128GB RAM
- 1× MIG/MPS-capable GPU (optional)
/HIGH UTILIZATION
Pro
- 24–64 CPU cores
- 256GB+ RAM
- 2–4 GPUs with MIG/MPS, NVMe scratch
/OPERATOR SCALE
Fleet
- Multi-node scheduling
- Redundant networking
- Standardized images + monitoring
MIG/MPS and why it matters
Fractional GPUs only work if isolation and concurrency are real. MIG partitions GPU resources; MPS improves multi-process utilization. Kova’s scheduling assumes these primitives when you want predictable $/result.
“Certification” is represented as a waitlist CTA in this demo build.