Performance
Same block. Every site. Measured before handover.
We publish measured results when the first block goes live in Q2 2027. Until then, this page shows how we will measure.
How we measure
Three numbers, measured the same way at every site.
- 01
Latency from the US
Round trip time from US hubs to the block, measured continuously and reported per contract.
- 02
Fabric bandwidth
All reduce bandwidth across the full block with NCCL tests, before handover and after every maintenance window.
- 03
Acceptance
Every node passes GPU stress, memory and thermal burn in before a cluster is handed over. Results ship with your handover pack.
Results
Published at go live, not before.
| Metric | Method | Result |
|---|---|---|
| RTT to Central US hub | Continuous probes, p50 and p99 | Q2 2027 |
| RTT to US South hub | Continuous probes, p50 and p99 | Q2 2027 |
| Block all reduce bandwidth | NCCL tests, 256 GPUs | Q2 2027 |
| Burn in pass rate | GPU, memory and thermal stress | Q2 2027 |
Questions about the method? We will plan a benchmark on your workload.
Tell us the model, the scale and where your users are. We size a test and agree the metrics before you sign.