DC Fabric Architecture
Scalable leaf-spine designs aligned with GPU workload behaviour, east-west traffic patterns, and long-term growth requirements.
Data Center Modernization
& AI Fabrics
PalC designs and delivers open, AI-ready data center fabrics for high-throughput GPU communication, predictable latency, and operational stability across training and inference environments.
KEY CAPABILITIES
Built for AI workloads from day one
Comprehensive fabric engineering -- from leaf-spine design to storage access and operational telemetry.
AI-Aware Network Design
Network designs engineered for the east-west traffic, high data movement, and performance sensitivity of training and inference platforms.
Open & Disaggregated Networking
SONiC-based open networking on multi-vendor hardware -- full operational control, zero proprietary lock-in.
High-Performance Storage Access
NVMe-oF over RoCE transport for low-latency, high-IOPS tiered storage architectures for AI workloads.
Observability-First Design
Telemetry, monitoring, and diagnostics embedded at fabric design time -- gNMI streaming, Grafana dashboards, and real-time visibility from day zero.
Production Validation & Readiness
IntelliSuite-driven validation against real traffic, scale limits, and failure scenarios before any production cutover.
Modular, scalable fabric architecture
Leaf-spine at the core, RoCE transport for GPU communication, NVMe-oF for storage, and full observability from Day 0.
Click a component in the diagram or panel to explore details.
Spine Switch 1100GbE/400GbE spine
Spine Switch 2100GbE/400GbE spine
Leaf Switch 125GbE/100GbE leaf
Leaf Switch 225GbE/100GbE leaf
Leaf Switch 325GbE/100GbE leaf
GPU PodNVIDIA A100/H100 cluster
NVMe-oF ClusterHigh-speed storage fabric
Observability LayerPrometheus, Grafana, ELK
DPU OffloadSmartNIC/DPU acceleration
RoCE PipelineRDMA over Converged Ethernet
Components
SONiC Open Networking
Open-source network operating system for disaggregated infrastructure.
- Vendor-agnostic switches with standardized NOS
- 40–60% cost reduction vs. proprietary
- Full control for AI workload customization
- Scales without lock-in
Used across spine, leaf, and fabric switches.
RoCE Optimization
RDMA over Converged Ethernet for lossless GPU-to-GPU communication.
- Sub-microsecond inter-GPU latency
- Zero packet loss for AI traffic
- 30–50% faster model training
- PFC, ECN, and buffer tuning
Critical for GPU pods and AI clusters.
Storage & Data Management
High-performance storage fabric with NVMe-oF for AI workloads.
- Direct GPU-to-storage, minimal latency
- Tiered hot/cold for cost and performance
- High-throughput for training pipelines
- NVMe-oF over TCP and RDMA
Used across GPU pods, NVMe-oF clusters, and fabric switches.
Accelerated AI Networking
High-performance network fabrics optimized for AI/ML workloads.
- 100GbE and 400GbE leaf-spine
- Predictable latency for distributed training
- Efficient GPU cluster communication
- Intelligent traffic engineering
Spans GPU pods, DPU offload, and fabric.
Scalable & Open Network Fabrics
Flexible, vendor-neutral network architectures that scale with your needs.
- Scale from pilot to multi-rack without redesign
- EVPN-VXLAN for secure multi-tenant
- Automation-first operations
- Multi-vendor, no lock-in
Used across spine and leaf layers.
End-to-End AI Infrastructure Monitoring
Real-time observability across compute, network, and storage.
- GPU utilization & job health
- Network latency & congestion
- Storage IOPS & throughput
- Proactive issue detection
Used across GPU pods, NVMe-oF clusters, and fabric switches.
Network Speeds
- Spine Interconnect400GbE
- Leaf Uplinks100 / 200GbE
- Server / GPU NIC100 / 200GbE
Protocols
- GPU TransportRoCEv2
- Congestion CtrlPFC · ECN · DCQCN
- OverlayEVPN-VXLAN
Storage
- ProtocolNVMe-oF (TCP/RDMA)
- ArchitectureTiered · Distributed
- Per Pod IOPS>1 Million
GPU Support
- GPU ClassNVIDIA A100 / H100
- Intra-nodeNVLink
- Inter-nodeRoCE v2
TECHNICAL CONFIGURATION
RoCE + NVMe-oF — configured for production
Lossless RoCE transport and high-performance NVMe-oF storage, tuned to meet AI traffic patterns and production-scale workloads.
SONiC — PRIORITY FLOW CONTROL
PFC for Lossless RoCE Transport
Zero packet loss for RoCE transport across configured queues and lossless buffer pools.
SONiC PFC configuration for RoCE
{
"PORT_QOS_MAP": {
"Ethernet0": {
"pfc_enable": "3,4"
}
},
"BUFFER_POOL": {
"ingress_lossless_pool": {
"type": "ingress",
"size": "139458560",
"xoff": "20971520"
}
}
}| PFC Queues | 3 & 4 (RoCE) |
|---|---|
| ECN Mark | DCQCN |
| Packet Loss | Zero |
| GPU Latency | <1µs |
KUBERNETES — NVMe-oF STORAGECLASS
NVMe-oF for GPU-pod Storage
High-IOPS distributed storage tuned for AI workloads with striped pools.
Kubernetes StorageClass for AI workloads
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
name: nvmeof
provisioner: nvmeof.csi.openebs.io
parameters:
replicas: "3"
poolType: "striped"
allowVolumeExpansion: true| Protocol | RDMA / TCP |
|---|---|
| IOPS per Pod | >1 Million |
| Replicas / Stripe | 3-way stripe |
| Latency Class | Memory-class |
PERFORMANCE
SLOs, Latency & Scale Benchmarks
Designed for AI fabrics, cloud interconnects, and enterprise data centres under real production load.
<1µs
Inter-GPU Latency
Network latency between GPU nodes via RoCE v2 lossless fabric.
>80%
Bandwidth Utilisation
Sustained during large model training workloads at full scale.
>1M
Storage IOPS
Per GPU pod via NVMe-oF with RDMA transport and tiered storage.
>90%
GPU Utilisation
Achieved during training — network no longer the bottleneck.
USE CASES
Who deploys this solution
Purpose-designed infrastructure for the environments where AI performance, reliability, and scale are non-negotiable.
AI & Machine Learning Platforms
Data center networks supporting large-scale distributed training, inference pipelines, and GPU-dense environments—where network latency directly impacts model iteration speed.
BFSI & Regulated Enterprises
Highly available, observable data center networks for transaction systems, analytics platforms, and compliance-driven environments—where audit trails and predictable performance are required.
Cloud-Adjacent & Hybrid Data Centers
Modern data center networks designed to integrate cleanly with public cloud—consistent networking policies, automation-first design, and DC interconnect for hybrid workload placement.
High-Growth Digital Platforms
Environments where rapid scale and frequent infrastructure change demand stable, predictable network behavior—open disaggregated architectures that grow without re-architecture.
HOW WE WORK
Design → Build → Validate → Operate
A proven engineering methodology that delivers production-grade results with operational excellence from day one.
STEP 1
Design
Business goals, workload profiles, and scale requirements translated into architecture and fabric designs.
STEP 2
Build
Engineering open fabric configurations, integration, and deployment tooling for production environments.
STEP 3
Validate
IntelliSuite-driven testing against real traffic, scale limits, and failure scenarios before cutover.
STEP 4
Operate
Ongoing support, telemetry monitoring, and continuous optimization so the fabric remains healthy long term.
INTEGRATED CAPABILITIES
SONiC Integration
Cloud Integration
IaC/GitOps
Observability
RoCE Optimisation
NVMe-oF Storage
Faster AI Readiness
Deploy and scale AI workloads without infrastructure bottlenecks.
Predictable Performance
Consistent latency and throughput under varying load conditions.
Full Visibility
Real-time telemetry for proactive issue detection and resolution.
No Vendor Lock-in
Open architectures and multi-vendor hardware preserve long-term flexibility.
Scales Operationally
Observability-first design keeps operations manageable as networks grow.
Proven outcomes from the field
Deployments across AI fabrics, multi-cloud, automation, and security.
ODM PARTNERS
TRUSTED BY LEADING TECHNOLOGY PARTNERS
Next steps
Planning your AI-ready data center fabric?
Share your SLO targets for latency, bandwidth utilization, storage IOPS, and visibility. PalC will help design an open, production-ready RoCE + NVMe-oF architecture.
