A building-permits computer-vision farm must harden GPU nodes so only approved inference frontends can reach them. Which network security evaluation best fits that goal?

Building-permits computer-vision farms hold high-value GPU nodes that should accept traffic only from approved inference frontends. Enabling DHCP snooping with dynamic ARP inspection on inference frontend VLANs validates IP-to-MAC bindings for client sessions, but it does not enforce which approved inference frontends may reach GPU node interfaces; address integrity is not an access-control allowlist, and RoCE lossless behavior does not remove the need for reachability filtering. Hardening preserves RoCE/training performance while still enforcing reachability policy at leaf or policy-engine layers.

Deploying MACsec encryption on spine-leaf links secures RoCE transport between GPU nodes and approved inference frontends, but it does not restrict which frontends can reach GPU node interfaces; link-layer confidentiality is not access control and leaves the frontend allowlist requirement unmet. Isolating GPU management interfaces in a dedicated management VRF with authenticated jump-host access controls platform operations for patching and monitoring, yet it does not govern whether approved inference frontends can reach GPU data interfaces; it addresses admin-plane access, not the frontend-to-GPU reachability policy. Dismissing ACLs because lossless fabrics supposedly cannot coexist with security controls is a false dichotomy, and replacing them with DHCP snooping or dynamic ARP inspection is equally insufficient: neither provides an allowlist of approved inference frontends at the GPU node interface, and both fail to enforce the required reachability control.

Exam caveat: overly broad deny ACLs can break legitimate health checks, orchestration agents, and storage mounts—policy must whitelist the full approved frontend and control-plane set, not only the user-facing API VIP. Operational check: from an untrusted municipal VLAN, attempt reachability to GPU data/management IPs and confirm deny; from approved frontends, confirm allow; review ACL counters and change tickets so permit-farm hardening stays aligned after frontend IP changes.