DPU offload
BlueField flow agent
An OpenFlow 1.3 and gRPC agent that compiles rules into mlx5 hardware-steering template tables, so every rule is in hardware before the first packet.
- DPDK
- rte_flow
- BlueField
AI infrastructure engineering
Build AI on solid ground. Then climb.
StrataPath designs and builds the layers AI workloads run on: GPU fabrics, DPU and SmartNIC offload, kernel-bypass datapaths and eBPF security. We find where the time goes, then move the work to where it runs best.
Why AI clusters underperform
The GPUs are rarely the problem. The layers underneath are.
A fabric that fights the job
Congestion, pause storms and miscabled rails turn every all-reduce into the slowest step of a training run.
Host CPUs doing NIC work
Classification, NAT, encryption and firewalling burn cores that a DPU or NIC eSwitch could handle at line rate.
Security bolted on late
Model weights, secrets and egress need enforcement in the kernel, not just a policy document.
What we do
Three layers of work, built from the ground up.
03 Workloads
02 Datapath
01 Fabric
Secure and tune
eBPF and XDP enforcement, telemetry and performance tuning for the workloads on top.
Offload and accelerate
DPDK fast paths and BlueField / ConnectX offload with rte_flow, OVS-DOCA, SR-IOV and virtio-net.
Fabric and transport
Rail-optimized RoCEv2 fabrics, collective transport for NCCL and multipath RDMA, and the configs to run them.
How we work
A clear path, one step up at a time.
-
Profile
Measure where packets and cycles go: PCIe, NIC queues, cores, memory and fabric counters.
[Duration]
-
Design
Write it down: where every stage runs, what gets offloaded, and how each failure is handled.
[Duration]
-
Build
Ship working code in C, C++, Go or Rust, with tests and a mock backend that runs in CI.
[Duration]
-
Benchmark
Prove it against a baseline (Mpps, latency percentiles, jitter), then hand over ownership.
[Duration]
Selected work
Engineering we can show you.
GPU fabric
RoCEv2 fabric generator
One JSON spec becomes cabling, FRR, QoS and per-host NCCL configs for a rail-optimized fabric. Python, C and Go versions give byte-identical output.
- RoCEv2
- NCCL
- Go / C / Python
AI workload security
aiguard
eBPF container security for AI workloads: 22 BPF LSM programs enforce exec, model-file, secret and egress allowlists, with a learning mode.
- eBPF
- BPF LSM
- libbpf
[A real client quote about what working with StrataPath changed.]
Find the bottleneck before you buy more hardware.
Bring a workload and a target. In one call, we'll help you work out where the time goes and which layer to fix first.
Book a technical call