Network engineering, optical expertise and supply-chain knowledge. We design, source and integrate infrastructure for carriers, datacenters and AI.
Engineering
The fabric, optical path, routing policy and host configuration have to work together. Our engineering spans the design choices, integration details and measurements that determine how the installed network behaves.
How we structure deliveryAccelerated computing
Training, distributed inference and storage place different demands on a network. We design AI/ML fabrics from the communication pattern outward, then tune the hosts, transport and switches as one system.
Map collective communication, job placement and storage traffic to the fabric. Evaluate rail alignment, shared versus dedicated traffic networks, and single- or multi-plane Clos topologies.
Size switch radix, bisection bandwidth and oversubscription for the job scale. Account for all-to-all traffic, incast, expansion and the capacity that remains during a failure.
Align DSCP or PCP classification, traffic classes, queues and scheduling across NICs and switches. Coordinate ECN marking, CNP feedback and endpoint rate control such as DCQCN where supported.
Select lossless operation with PFC and ECN, or an ECN-based lossy design supported by the platform. Tune thresholds and rate control against the actual ASIC, NIC and workload.
Where PFC is used, confine it to the required priorities. Set XOFF/XON thresholds and buffer headroom for the ASIC, line rate, MTU, cable length and pause-response delay.
Check pause propagation, head-of-line blocking and deadlock recovery. Use congestion feedback to control load before sustained pausing spreads through the fabric.
Test ECMP hashing, flow entropy and collisions against collective traffic. Evaluate adaptive routing or packet spraying only when the switch and transport support them.
Measure path imbalance, reordering and recovery during link failures. Balance steady-state throughput against the behavior of the surviving fabric.
Resolve PCIe topology, GPU-to-NIC affinity, NUMA placement and GPUDirect RDMA prerequisites. Check peer access, ACS/IOMMU behavior and firmware, driver and communication-library compatibility.
Align MTU and traffic marking end to end. Account for topology visibility in containers or VMs, and the interaction between backend, storage, frontend and management networks.
Evaluate 400G, 800G and 1.6T interfaces against the exact switch, NIC and module support. Match electrical lanes, breakout modes, FEC, thermal limits and management behavior.
Verify optical levels, fiber polarity, connector cleanliness and loss. Correlate lane errors, uncorrectable FEC events and link flaps with stalled communication.
Workload performance. RDMA throughput and latency, plus NCCL or equivalent collectives across representative message sizes, scales and placements.
Congestion and resilience. Collective completion time alongside queue occupancy, ECN/CNP activity, PFC duration and physical errors, including competing traffic and failures.
Optical transport
A coherent interface’s nominal reach is only a starting point. We engineer the operating mode, fiber route and line system together, with a power and noise budget that can be checked against the installed path.
Match line rate, baud rate, modulation, shaping and FEC at both ends. Check the required host software, module power and thermal envelope.
Evaluate 400ZR, ZR+ and transponder options by their exact interoperable modes. Similar form factors or marketing labels do not establish compatibility.
Build span-loss and received-power budgets alongside the selected mode’s OSNR or GSNR requirements. Include connectors, splices, amplifier noise and nonlinear penalties.
Check receiver sensitivity and overload separately from signal quality. Reserve operating margin for aging, repairs and changes in channel loading.
Set launch power, gain, tilt and equalization across the route. Evaluate EDFA and Raman amplification where appropriate to the fiber and line system.
Model partial and full loading, wavelength additions and amplifier failures. More optical power can increase nonlinear penalties rather than improve margin.
Match occupied bandwidth and channel spacing to mux and ROADM passbands. Account for cascaded filtering and the actual route through the system.
Check alien-wavelength compatibility, power targets and monitoring capabilities before introducing a new coherent interface into an existing line system.
Account for chromatic dispersion, polarization-mode dispersion, attenuation and the receiver’s compensation limits.
Translate shared spans, amplifiers and site dependencies into protection requirements. Verify that the proposed restoration path has sufficient capacity and optical margin.
Establish transmit/receive power, pre- and post-FEC behavior, uncorrectable codewords and the available Q, Q-margin or SNR measurements for the active mode.
Use OSNR measurements where applicable and understood. Record alarm thresholds and a stable baseline so degraded margin is visible before a hard failure.
Optical performance. The exact endpoint and line-system combination, measured against its engineered power/noise budget and acceptance limits.
Operational margin. Stability under representative channel loading, maintenance and protection events, with channel plans and measurements retained for future changes.
Packet networks
We design packet networks from datacenter underlays to carrier backbones and the Internet edge. Policy, forwarding scale and convergence are engineered together so the control plane and the traffic path behave as intended.
Define IPv4/IPv6 transit, peering and customer policy with explicit import/export rules, communities and traffic preferences.
Combine IRR-based filtering, RPKI origin validation, maximum-prefix limits and route-leak controls. Verify normal updates, exceptional routes and the consequences of policy changes.
Choose BGP/ECMP underlays, EVPN/VXLAN overlays and MPLS services to match scale, isolation and the operating model.
Check route, MAC and neighbor capacity, encapsulation overhead and end-to-end MTU. Keep service boundaries and failure domains explicit.
Assess capacity, latency and path control using actual traffic. Coordinate QoS, failure detection, routing convergence and forwarding-table updates.
Test partial failures, asymmetric paths and maintenance drains as well as complete outages. Aggressive BFD timers must fit the platform’s scale and processing limits.
Derive addressing, policy and configuration from maintained source data. Use repeatable validation, staged changes and a defined rollback path.
Compare routing intent with installed routes and forwarding entries. Correlate telemetry, flow data, queueing and drops to locate problems across the path.
Policy and forwarding. Representative routes, filtering, reachability, isolation and forwarding consistency across the intended services.
Failure behavior. Convergence time, loss, latency and surviving capacity during peer/link failures, withdrawals and maintenance.
Integration and delivery
Deployment joins equipment, software, cabling and operating practice. We turn the design and bill of materials into an integration sequence, then verify the installed system and document what the operating team needs to run it.
Integrate switches, routers, compute hosts and storage using the intended firmware, drivers, operating systems and interface modes.
Investigate bottlenecks across PCIe, CPU/NUMA, storage I/O and the network. Resolve compatibility and licensing dependencies before the deployment window.
Account for power feeds, cooling, airflow, rack density and service access alongside the hardware and topology.
Define cable paths, patching, port maps, fiber polarity, labeling and spares. Check the physical installation against the logical design before turn-up.
Build configuration, addressing, routing, monitoring and management-access baselines. Coordinate hardware arrival with circuits, cross-connects and site readiness.
Sequence migrations into testable stages with prerequisites, acceptance criteria and rollback points. Validate representative workloads before committing the next stage.
Document the installed topology, inventory, configurations, measurements and known limits. Make operational ownership and escalation paths explicit.
Test console and out-of-band access, configuration restoration and recovery with the production network unavailable. A saved backup must be usable when it is needed.
Installed system. Physical connectivity, workload performance, failover, monitoring and recovery across the complete path.
Handover. As-built records, acceptance results and recovery procedures, including unresolved constraints and the capacity available for expansion.
Supply chain & procurement
We source constrained optical transceivers and network hardware, negotiate pricing and check alternatives against the deployment requirements. Manufacturer access, technical qualification and delivered cost all shape the purchasing decision.
We identify workable sources for high-speed transceivers and network hardware using our knowledge of manufacturers, distribution channels and component availability.
We check manufacturer and part identity, exact configuration, quantities and delivery requirements with each source, and distinguish confirmed availability from indicative lead times and proposed alternatives.
We evaluate how volume, sourcing route, lead time and support terms affect price. We negotiate with the component requirements and market conditions in view.
Delivered cost, freight, payment terms and warranty coverage all belong in the comparison. We put those details alongside the technical specification so you can evaluate complete offers.
We assess substitutes against host support, electrical and optical interfaces, FEC, power, temperature and fiber requirements.
We define the qualification and approval needed before an alternative enters the bill of materials. Vendor coding or a matching headline data rate does not establish system compatibility.
We build a coordinated sourcing plan for switches, routers, optics, fiber, licenses and spares, with quantities and delivery phases tied to installation windows.
We track long-lead dependencies, substitutions and expansion needs, keeping purchasing decisions connected to the configuration that engineering has qualified.
Optical transceivers are a particular strength. We work through the required interface, qualify alternatives, identify available supply and compare the complete delivered cost.
Discuss your sourcing needsStrategy & insight
We assess interconnection, facilities, fiber routes and technology choices against traffic demand, capacity, operating cost and expansion plans.
We evaluate transit, public exchanges and private peering against traffic destinations, volume, latency and reach.
Ports, transport, cross-connects and operational overhead all affect the economics. We connect the commercial choice to routing policy, capacity thresholds and resilience.
We compare sites by carrier access, power, cooling, density and room to grow, alongside the datacenter interconnect and replication needs of the compute footprint.
We map circuit delivery, cross-connect availability and migration dependencies before committing to a location or schedule.
We assess switches, optical interfaces and AI-fabric options against workload requirements, software maturity and the installed environment.
Lifecycle cost, support, power and migration effort inform the recommendation: adopt a new platform, qualify an alternative or use existing infrastructure for the next phase.
We examine available route records, conduit, building entrances and carrier handoffs to identify common failure points. Separate providers do not necessarily provide physical diversity.
We distinguish confirmed routes from assumptions and outstanding surveys, then assess build-versus-buy options, construction dependencies and restoration requirements.
We develop a phased plan with investment priorities, capacity triggers, dependencies and decision points. We document the assumptions and evidence behind each choice so the plan can change as demand, technology and supply conditions evolve.
Discuss your next phaseProject delivery
A deployment connects engineering decisions to purchase commitments and operating responsibility. We structure the work around an agreed scope: what Petabit delivers, what your team and other suppliers own, and how the result will be accepted.
Establish the workload, topology, capacity, failure behavior and physical constraints. Set the deliverables, interfaces between teams and acceptance criteria before equipment choices become purchase commitments.
Connect exact hardware, firmware, interface modes and cabling to the design. Identify the evidence needed to qualify alternatives, alongside source availability, commercial terms, warranty coverage and spares.
Tie quantities and delivery phases to site readiness and installation windows. Track long-lead dependencies and configuration changes. Use testable stages, with clear prerequisites and rollback points, to carry the plan into service.
Evaluate the installed system against the agreed tests. Record the configuration, test conditions, results and unresolved exceptions. Bring as-built records, recovery procedures, ownership and escalation paths into the handover.
Engineering, component supply and integration can be scoped individually or together. The agreed scope sets the deliverables, acceptance process and support responsibilities for that engagement.
Discuss a deployment
Petabit Scale / WhiteBox Networks
WhiteBox Networks is Petabit Scale’s hardware brand for optical transceivers, fiber assemblies and cabled interconnects.
Explore WhiteBox Networks
