Mandarum Voltra — Pack every GPU, hold every SLO.
An agentic scheduler and autoscaler that places each inference request across a mixed NVIDIA GPU fleet — batching, routing, and packing with NVIDIA Dynamo disaggregated serving and MIG partitioning. It learns from served traffic and DCGM telemetry to raise utilization from ~35% to 80%+ without missing p99.
Why GPU fleets waste half their budget.
Fleets sit at 30–40% utilization
Static schedulers over-provision to be safe and waste GPUs. Voltra continuously re-packs models and requests across the fleet to keep expensive accelerators busy.
Bursty traffic blows p99
Round-robin routing ignores batch shape and live load. Voltra places prefill and decode separately via Dynamo and autoscales each phase to hold first-token latency.
No one routes by price
The same request can run on a cheaper-capable GPU. Voltra routes per request to the lowest-cost GPU that meets the SLO and caps spend at the scheduler.
What Voltra does.
Agentic placement across the fleet.
A reinforcement-learning policy ingests live DCGM telemetry and request shape, then batches, routes, and packs each call onto the right GPU and MIG slice. The policy improves from your own served traffic.
- → RL policy trained on served traffic + DCGM
- → MIG / MPS-aware bin-packing
- → Per-request routing to cheapest-capable GPU
await voltra.place({ model: "meta/llama-3.1-70b", slo: "p99<200ms", fleet: "h100,h200,l40s" }); // packed @ 81% util
Disaggregated autoscaling via Dynamo.
Voltra separates prefill and decode with NVIDIA Dynamo and autoscales each phase independently, maximizing NVLink bandwidth on HGX clusters and absorbing bursts without cold-start latency.
- → Prefill / decode disaggregation via Dynamo
- → Independent autoscaling per phase
- → NVLink-aware placement on HGX
Test policies before they hit prod.
A built-in simulator replays your fleet from DCGM traces so you can validate a scheduling policy against real load and predict utilization, p99, and $/token before rollout.
- → Policy simulator on real traces
- → Utilization + p99 + cost forecast
- → Safe staged rollout with instant rollback
Voltra, in production.
One product. One platform.
Voltra routes cheaper. Praxon governs what it routes. Cuvex grounds what it serves. Edgeon carries the same runtime and scheduling to your Jetson factory floor. Twinex lets you simulate a Voltra policy before it ever touches prod. Start with scheduling — grow into the OS.
Mandarum Praxon
Run autonomous agents with per-action identity, NeMo Guardrails, durable execution, and an immutable audit trail.
Explore Praxon →
Mandarum Cuvex
GPU embedding, cuVS vector search, and reranking in one permission-aware recall path.
Explore Cuvex →
Mandarum Edgeon
Serve governed AI on NVIDIA Jetson fleets with cloud-parity policy, telemetry, and safe OTA.
Explore Edgeon →
Mandarum Twinex
Simulate GPU inference fleets on Omniverse to plan cost, capacity, and serving topology before production.
Explore Twinex →Stop leaving money in idle GPUs.
See how Voltra packs your fleet, holds your SLOs, and cuts $/token from the first request. Start free, or let the team walk you through a fleet simulation on your own hardware.