Technology

The Stack Under the Hood

RISC-V silicon, programmable eBPF hardware VMs, and hardware-enforced Trusted Execution Environments.

RISC-V ISA

RISC-V Packet Processing

The open RISC-V ISA enables PacketFive to define custom extensions specifically optimised for packet parsing, header manipulation, and flow-table lookups — operations that consume disproportionate CPU cycles on conventional architectures.

  • Custom vector and packet-processing ISA extensions
  • Deterministic single-cycle header operations
  • Open-source toolchain (GCC/LLVM) fully supported
  • Designed for high-bandwidth AI/HPC fabric interconnects

Why RISC-V?

Proprietary ISAs lock customers into vendor ecosystems and prevent custom extensions. RISC-V's open, modular specification lets PacketFive add domain-specific instructions without licence fees or NDA constraints — and contribute improvements back to the community.

8× Parallel eBPF VMs

Each hardware eBPF VM is a full implementation of the eBPF ISA in silicon — not a software JIT compiler, but hardened gates running eBPF bytecode at line rate. Programs can be loaded, verified, and hot-swapped without data-plane downtime.

eBPF Hardware

Programmable at Wire Speed

eBPF has become the de-facto standard for programmable packet processing in Linux. KESTREL-V takes this a step further: instead of running eBPF in software, the ISA is implemented directly in RISC-V hardware VMs.

  • Full eBPF ISA implemented in silicon
  • Hot-swap programs with zero data-plane downtime
  • eBPF verifier enforces safety before execution
  • Compatible with standard BPF toolchains (libbpf, bpftool)
  • 8 independent VMs for parallel pipeline stages

Trusted Execution Environment

Hardware-enforced isolation ensures that sensitive network functions cannot be tampered with, even by privileged host software.

🔒

Attestation

Remote attestation proves to a verifier that the correct, unmodified eBPF program is running inside the enclave.

🛡

Isolation

TEE enclaves are hardware-isolated from the host OS and hypervisor — critical for multi-tenant AI cluster deployments.

🔑

Sealed Secrets

Cryptographic keys and policy material are sealed to the enclave, inaccessible to privileged host processes or hypervisors.

Built for AI/HPC Fabrics

GPU clusters for CUDA and ROCm workloads demand ultra-low-latency, high-bandwidth networking. KESTREL-V accelerates RDMA, RoCE, and MPI collective operations directly in the data plane.

🚀

RDMA / RoCE

Hardware-accelerated RDMA over Converged Ethernet for GPU-to-GPU memory transfers.

🔄

MPI Collectives

In-network computing for AllReduce, Broadcast, and Scatter-Gather — offloading collective operations from CPU/GPU.

📊

Telemetry

Per-flow telemetry and congestion signals streamed to HiCAIN for real-time cluster health monitoring.