Specifications
Product Family
Arista 7800R3 Series modular data centre switch router
Chassis Models
DCS-7804 (10RU), DCS-7808 (16RU), DCS-7812, DCS-7816 and DCS-7816L (32RU)
Fabric Capacity
230 Tbps (460 Tbps full duplex) system throughput
Per-Slot Bandwidth
Up to 14.4 Tbps per line card slot
Port Density
Up to 576 wire-speed 400G ports, or 768 x 100G ports
Port Speeds
100G, 200G and 400G with 800G-ready architecture
Interfaces
Standards-based OSFP and QSFP-DD 400G plus QSFP 100G
Forwarding Rate
Up to 96 billion packets per second
Latency
Under 4 microseconds for 64-byte packets
Packet Buffer
Up to 24 GB per line card
Fabric Architecture
Cell-sprayed fully scheduled fabric with virtual output queuing and distributed credit scheduling
Lossless Design
No head-of-line blocking, no noisy-neighbour effects, deep buffering for incast
AI Load Balancing
Cluster Load Balancing (CLB) operates on RDMA queue pairs with a global leaf and spine view
Scale
Two-tier topology scaling beyond 150,000 endpoints
Coherent Optics
400G ZR and ZR+ support on line cards
Security
MACsec, IPsec and VXLANsec encryption
Timing
IEEE 1588 PTP and next-generation timing services
Monitoring
LANZ microburst detection and accelerated sFlow (RFC3176)
Control Plane
Dual supervisors, multi-core hyper-threaded x86, 64 GB DRAM, 4 GB flash, 1+1 redundancy
Redundancy
Fabric module redundancy, N+N configurable PSU redundancy, hot-swappable components, front-to-rear airflow
Power Efficiency
16 W per 100G and 25 W per 400G port typical
Software
Arista EOS single image, CloudVision, Zero Touch Provisioning, eAPI, NETCONF and JSON-RPC
Overview
Training and inference clusters put a different requirement on the network than a general-purpose data centre: the back-end fabric must stay lossless and deterministic at sustained full load, because a single dropped packet can stall a collective operation across thousands of GPUs. The Arista 7800R3 Series addresses that with a fully scheduled cell-based fabric, virtual output queuing per port and deep buffering up to 24 GB per line card.
The platform is accelerator-agnostic and scales to more than 150,000 endpoints in a two-tier leaf-spine topology. Its Cluster Load Balancing feature is the AI-specific addition: it works on RDMA queue pairs rather than packet hashes, using a global view to optimise leaf-to-spine and spine-to-leaf flows simultaneously so large, few-flow training jobs complete sooner. The chassis range from 4 to 16 line-card slots lets a builder start small and expand fabric capacity in place.
Key Benefits
Lossless scheduled fabric keeps collective operations from stalling under 100% load. RDMA-aware Cluster Load Balancing shortens job completion time for large-bandwidth AI flows. Up to 576 x 400G ports and 230 Tbps in one chassis reduces spine count and cabling, and a single EOS image across the fleet simplifies automation through CloudVision and eAPI.
Applications
GPU training cluster back-end fabrics, AI inference fabric aggregation, HPC interconnects, storage and front-end networks inside AI halls, and large-scale leaf-spine builds that need accelerator-agnostic 400G spine capacity.
Request a Quote — ARISTA 7800R3 SERIES — 400G AI SPINE FOR GPU CLUSTERS
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →