Technology
The engine under every projection.
A packet-fidelity simulator and emulation platform, calibrated against measured runs and cross-checked against hardware emulation. This page is what “calibrated” means in practice.
What the substrate is.
- Packet-fidelity collective simulation
- Arbitrary-topology ingestion
- Packet-level simulation reach past 64k nodes
- Replay-validated against measured step times from captured traces
- Cross-checked against hardware emulation
What it does, and why each part matters.
Packet-fidelity collective simulation
Collectives simulated at packet level, not approximated by an analytical model of what a collective ought to cost.
Arbitrary-topology ingestion
Your topology, not a template. Fat-tree, dragonfly, rail-optimized, or whatever you actually built.
Device-mode statistics
Queue depth, PFC, ECN and CNP, RTT, and buffer occupancy — the signals that explain why a collective is slow rather than merely reporting that it is.
Memoized sweep harness
Design-space sweeps that reuse prior work instead of re-simulating it, which is what makes a ranked set of variants affordable rather than a single point answer.
Chakra as the input format
Workloads arrive as Chakra execution traces — the same format the standards work is converging on, so what you capture is not locked to us.
Topology reconstruction
The fabric you intended and the fabric your workload actually experiences are frequently different. That difference is a finding, not an error to be smoothed over.
How it runs
- Dedicated secure tenancy
- Elastic provisioning
- User, workspace and cost administration
The validation ladder.
Three rungs, each backed by different evidence. Reach and validation are separate claims and are stated separately here.
- 01
Replay validation
The simulator reproduces measured step times from captured traces. This is the gate: fail it and no projection is issued.
- 02
Hardware cross-check
Simulated behavior is cross-checked against hardware emulation, independently of the replay.
- 03
Beyond measurement
Past the point where either check reaches, the projection continues with a confidence band that widens with extrapolation distance — and says so on every number.
Packet-level simulation reach past 64k nodes describes how far the simulator runs. It is not a statement that results at that size have been validated against hardware — those are separate claims and we do not merge them.
Talk to an engineer.
Bring your topology, your transport, and the scale you care about. We will tell you what the simulator reproduces for that environment, what it does not, and where the confidence bands start to widen.