Arista announces first 3.2T switch

Arista Networks has unveiled the new R4 product family, announcing a significant technological leap in the field of data centers and AI networks. The announcement is centered around the new “HyperPorts,” which represent a novelty with 3.2T. Competitors are just now making the jump to 800G or 1.6T. These ports are intended to form the backbone for the distributed AI clusters of hyperscalers and large enterprises.

In total, Arista has introduced three new model series of its R4 family: 7020R4, 7280R4, and the flagship 7800R4. The 7020R4 series comes in three model variants in a single rack unit with 48x10G RJ45 copper and 4x100G QSFP for the 7020R4-48TX-4QC, over 48x1/10/25G SFP(+/28) and 4x100G QSFP for the 7020R4-48Y-4QC, up to the 7020R4-48Y-8QC, which offers 48x1/10/25G SFP(+/28) and 8x100G QSFP. They are intended as leaf switches. All can handle over 2.5 million IPv4 and IPv6 routes.

sFlow is available on all variants for controlling data flows. The two smaller models offer 2GB of dynamic packet buffer, and the large model even 4GB, which is significantly more than the competition. Additionally, Precision Time Protocol (PTP) and Synchronous Ethernet are also available.

The 7280R4 comes in two rack units and three different models. The 7280R4-64QC-10PE offers 64x100G QSFP and 10x800G OSFP slots, as well as a 16GB packet buffer. The 7280R4-32PE and 7280R4-32DE each offer 32x800G slots, with the PE having OSFP slots and the DE having QSFP-DD. Additionally, both have 8 x 25G SFP28 and 2 x QSFP-LS.

All three models feature an 8-core x86 CPU, 64GB DRAM, and a 480GB SSD, with the small model enabling 16GB packet buffer and a maximum throughput of 14.4 Tbit/s, and the larger ones 25.6 Tbit/s respectively. Arista specifies a low latency of 3.8 us for all. These switches also support 2.5 million IPv4 and IPv6 routes in their standard configuration and up to 5 million in the special K-series. Furthermore, they also provide sFlow, Precision Time Protocol (PTP), Synchronous Ethernet, MPLS, segment routing, and EVPN-VXLAN.

The spearhead is formed by the 7800R4 series, which comes with a modular chassis. With up to 576x800G or 128x3.2T slots and a maximum full-duplex throughput of 460 Tbit/s or 920 Tbit/s for the entire chassis, it delivers impressive figures. Arista calls the 3.2T slots HyperPorts. With these, Arista aims to address the increasing demand for lossless, highly parallel communication in the training and inference environments of modern AI systems.

The higher bandwidth can drastically reduce the number of required links, transceivers, and cables, which in turn is intended to improve energy efficiency and operating expenses. Arista has not yet specified which transceivers or direct-attach cables (DACs) can be used to equip the 3.2T slots.

Each slot in the chassis can transmit up to 28.8 Tbit/s and has a 32GB packet buffer. Variants with 4, 8, 12, and 16 chassis slots are available. An 800G linecard has 36x800G OSFP slots, although only the 7800R4K-36PE supports encryption with IPsec, MACsec, and VXLANsec, while all support sFlow. Latency per linecard ranges between 2.8 and 3.8 us.

Packet buffering and Comprehensive Encryption

In addition to pure port performance, Arista also integrates dynamic packet buffering and comprehensive encryption functions from IPSec to MACsec and VXLANsec to ensure secure, but also high-performance and low-latency data transmission between data centers. Furthermore, they support the protocols MPLS, segment routing, and EVPN-VXLAN, which are common in large enterprise and provider networks, as well as PTP. The Supervisor Engines come with six- or eight-core x86 CPUs, 64GB RAM, and 256GB SSD.

The 7800R4 systems and associated linecards are already available; the 7020R4 series and the 7800R4 variant with HyperPort support are scheduled for the first quarter of 2026. Arista did not comment on pricing. With the 3.2T HyperPorts, Arista is likely to further solidify its position in the data center and AI networking environment.

(emw)