Sub-µs Latency
Cut-through switching, RoCE v2 zero-copy, and adaptive routing deliver sub-microsecond hop latency. Your NCCL all-reduces finish faster — your training runs shorter.
/ NETWORKING
High-bandwidth fabric built for AI
Tscale Networking delivers the fabric that AI workloads actually need — 200/400/800 Gbps Ethernet, InfiniBand NDR, RoCE v2, and intelligent routing. Sub-microsecond latency, zero packet loss, full bisection bandwidth.
Full Bisection
Non-blocking Clos topology at 200/400/800 Gbps. Every server can talk to every other server at full line rate — no oversubscription, no hotspots.
Secure By Default
Wire-speed MACsec encryption on every link. Private VPCs, BGP communities, and segmentation baked into the fabric — no overlay tax.
/ FABRIC OPTIONS
Pick the right interconnect for the job
Different AI workloads have different network profiles. Tscale lets you mix and match fabrics per workload — RDMA over Converged Ethernet for inference, InfiniBand for tightly-coupled training, standard Ethernet for everything else.
- 200/400/800 Gbps Ethernet — the universal fabric, with RoCE v2 for GPU-direct RDMA.
- InfiniBand NDR / HDR — sub-microsecond latency for the largest multi-GPU training jobs.
- NVLink / NVSwitch — 900 GB/s GPU-to-GPU bandwidth inside the chassis.
- Adaptive routing — packets re-route around congestion in real time, no dropped flows.
/ INTERCONNECT PERFORMANCE
Distributed training at line rate
When 64 GPUs run an all-reduce, every microsecond of fabric latency becomes minutes of job time. Tscale’s interconnect sustains 94% of peak throughput across the entire 64-GPU job — not just the average, the worst case.
-
94% peak throughput
Sustained across 64-GPU all-reduces, validated on real training workloads.
-
142 µs all-reduce
Across 64 GPUs on a 200G RoCE fabric. Faster means shorter training runs.
-
Zero packet loss
PFC + ECN on every port, tuned for lossless RDMA. No silent retransmits.
/ PLATFORM
Everything around the fabric
Networking is more than wires and switches. Tscale’s platform includes routing, security, observability, and automation — all built for AI workloads from day one.
Topology
- Clos (leaf-spine) fabric
- Non-blocking at all speeds
- Adaptive routing (DLB)
- Explicit congestion control
- Cut-through switching
Routing
- BGP communities
- Private VPCs
- Floating IPs & anycast
- Load balancers (L4/L7)
- Direct Connect to your DC
Security
- Wire-speed MACsec
- Wire-speed IPsec
- Microsegmentation
- DDoS protection
- Network policy as code
Observability
- Per-flow telemetry (sFlow)
- Streaming metrics
- RDMA counter dashboards
- Congestion heatmaps
- Alerting & PagerDuty
Interconnects
- NVLink / NVSwitch
- InfiniBand NDR / HDR
- RoCE v2 (RDMA over Ethernet)
- Standard TCP/IP
- Direct RDMA to storage
Automation
- Terraform provider
- BGP API (Radar)
- Auto-VPC provisioning
- Cable & patch management
- Capacity planning
Performance
Every port, every link
Non-blocking Clos fabric — no oversubscription, no per-rack limits, no surprise throttling.
With InfiniBand NDR
Sub-microsecond across the fabric, validated on real training jobs in production.
Lossless RDMA fabric
PFC + ECN end-to-end, tuned for NCCL and RCCL. No silent retransmits, no stalls.
Backed by enterprise SLA
Redundant spines, hitless upgrades, 24/7 SRE on call for fabric-wide incidents.
For teams that need predictable performance
Multi-region Peering
Private interconnects between Tscale regions with sub-10ms RTT. Run a multi-region training job without the public internet in the path.
Learn MoreDirect Connect
Dedicated 10/100/400 Gbps private lines from your data centre to Tscale. Predictable latency, no shared infrastructure, no surprises.
Learn MoreThe fabric is the foundation
Networking ties every Tscale service together. Pair it with compute, instances, and managed Slurm to run distributed training without bottlenecks.