CudaFlow registers your GPU nodes, dispatches simulation jobs, and meters usage — all from a single control plane built for deep-tech compute.
30×
faster than CPU offload solvers
<30s
offline detection latency
0
queue services required
One API, five agent endpoints, zero external queue services.
Agents register over a single HTTPS call with an enrollment token. GPU specs, driver version and compute capability detected automatically.
SELECT FOR UPDATE SKIP LOCKED in Postgres — no SQS, no Redis. Jobs are claimed by the fastest available node.
Node liveness is checked at read time against last_seen_at. No background sweepers, no false positives.
GPU-seconds are recorded per job with idempotency keys. Duplicate agent reports are silently acknowledged.
Multi-tenant from day one. Argon2id passwords, session tokens, org membership roles, and enrollment token revocation.
Claude Sonnet and Opus 5 via Amazon Bedrock — built into the control plane for job analysis and intelligent automation.
Sign up, create your organisation, and generate an enrollment token.
Run the agent on any CUDA-capable machine. It registers itself and starts polling for jobs.
POST a job via the API or dashboard. The next available node claims and runs it.
GPU-seconds are recorded automatically. Export usage for billing or analysis.