Arista 7060 Series Switches: Architecture, Models, and Deployment Guide (2026)

The Arista 7060 Series is one of the most extensively deployed fixed-configuration switching platforms in modern data center networks. Spanning five generations of Broadcom Tomahawk merchant silicon, the 7060 family has evolved from early 40GbE/100GbE leaf switches into high-density 400GbE and 800GbE platforms purpose-built for AI/ML training clusters, hyperscale cloud fabrics, and enterprise spine-leaf architectures.

This guide provides an engineer-focused deep dive into the Arista 7060 platform: its silicon evolution, architectural intent, per-family hardware capabilities, deployment patterns, and practical selection criteria. Whether you are designing a new EVPN-VXLAN fabric, scaling an existing leaf-spine topology, or building out GPU cluster interconnects, this resource will help you understand where each 7060 generation fits and how to choose the right model for your environment.

Table of Contents

• 1. Introduction: Why the Arista 7060 Series Remains a Data Center Foundation
• 2. Positioning the 7060 Within Arista’s Switching Portfolio
• 3. 7060 Series Generations and Evolution
• 4. Silicon Architecture Deep Dive: Tomahawk → TH+ → TH3 → TH4 → TH5
• 5. 7060CX and 7060CX-32S (First-Generation Tomahawk, 40G/100G)
• 6. 7060CX2 and 7060SX2 (Tomahawk+, 25G/100G Leaf)
• 7. 7060X4: DX4 and PX4 Families (Tomahawk 3, 400G)
• 8. 7060X5: DX5, PX5, and CX5 Families (Tomahawk 4, 400G/800G)
• 9. 7060X6 Family (Tomahawk 5, 800G AI/ML Platforms)
• 10. Latency, Buffering, and Dynamic Buffer Allocation
• 11. OSFP vs QSFP-DD: Choosing the Right Optics Form Factor
• 12. Common Deployment Architectures
• 13. EVPN-VXLAN Fabric Designs with 7060 Platforms
• 14. AI/ML and HPC Fabric Design
• 15. Storage, Microbursts, and East-West Traffic
• 16. Migration Paths: 10G → 25G → 100G → 400G → 800G
• 17. Model Selection Framework
• 18. Generational Comparison: Key Specs Across 7060 Families
• 19. 7060 Series vs Arista 7280 Series: Architectural Differences
• 20. Buying Used and Refurbished Arista 7060 Switches
• 21. Why Engineers Source 7060 Hardware from BrightStar Systems
• 22. Related Resources

1. Introduction: Why the Arista 7060 Series Remains a Data Center Foundation

The Arista 7060 series was designed from the outset to serve as a high-performance, low-latency leaf and top-of-rack (ToR) platform for modern data center fabrics. Rather than attempting to address every possible traffic profile, the 7060 focuses on three core strengths:

• Deterministic low latency for predictable east-west forwarding
• High port density in compact 1RU and 2RU form factors
• Consistent EOS software across every generation, from the original CX through the latest X6

This design philosophy has made the 7060 the default leaf switch in thousands of spine-leaf deployments worldwide. As server interface speeds have progressed from 10GbE through 25GbE, 100GbE, 400GbE, and now 800GbE, Arista has iterated the 7060 platform in lockstep — each generation introducing new Broadcom Tomahawk silicon while maintaining operational consistency through Arista EOS.

The result is a platform family that spans five silicon generations, supports interface speeds from 1GbE to 800GbE, and addresses use cases from traditional enterprise leaf-spine fabrics to cutting-edge AI/ML training cluster interconnects.

BrightStar Systems works extensively with all 7060 subfamilies and can help design leaf-spine architectures, AI fabric topologies, and upgrade paths across generations. For full inventory, see our Arista 7060 Series catalog.

2. Positioning the 7060 Within Arista’s Switching Portfolio

Understanding where the 7060 fits within Arista’s broader portfolio is essential for making sound design decisions. Each Arista platform family is optimized for a specific set of traffic profiles, buffering requirements, and deployment roles.

The 7060 series is built on Broadcom Tomahawk-class ASICs, which prioritize high throughput, dense port counts, and low cut-through latency. These are shallow-to-moderate buffer platforms designed primarily for predictable east-west traffic in leaf and ToR roles.

The 7280R series, by contrast, is built on Broadcom Jericho-class ASICs with deep VOQ (Virtual Output Queuing) buffers — hundreds of megabytes of packet memory designed to absorb microbursts and handle congested aggregation scenarios. The 7280R family excels at spine, aggregation, DCI, and storage-heavy roles where burst absorption is critical. For a comprehensive breakdown of the 7280R architecture, see our Arista 7280R Series architecture guide.

The 7050X series shares many characteristics with the 7060 but is optimized for different port density and speed profiles. 7050X platforms are commonly deployed in mixed 10G/25G/100G environments where SFP-based server access density is the primary requirement.

In practice, many large-scale data centers deploy 7060 switches at the leaf layer feeding traffic into 7280R platforms at the spine or aggregation layer — combining the 7060’s low-latency density with the 7280’s deep-buffer congestion management.

3. 7060 Series Generations and Evolution

The 7060 platform has evolved through five distinct hardware generations, each aligned with a new generation of Broadcom Tomahawk silicon and a corresponding step forward in interface speed, buffer capacity, and fabric scale.

Each generation maintains full backward compatibility with Arista EOS, meaning operational workflows, automation tooling, and monitoring infrastructure built for earlier 7060 models carry forward without modification. This operational consistency is a key reason why organizations continue to standardize on the 7060 platform as they scale from 10G to 800G.

4. Silicon Architecture Deep Dive: Tomahawk → TH+ → TH3 → TH4 → TH5

The performance characteristics of every 7060 model are defined by its underlying Broadcom Tomahawk ASIC. Understanding the silicon evolution is critical for matching the right platform to your workload requirements, growth trajectory, and fabric design.

4.1 Tomahawk (TH) — First-Generation 7060CX

The original Tomahawk ASIC powers the 7060CX-32S and related 7260CX models. This silicon introduced merchant-silicon 100GbE switching at scale, delivering:

• 6.4 Tbps aggregate switching capacity (32 × 100G)
• 16 MB shared packet buffer with Dynamic Buffer Allocation (DBA)
• Cut-through latency as low as 450 ns
• Flexible port breakout: every QSFP100 port supports 100G, 40G, 4×25G, or 4×10G modes

Tomahawk-based 7060CX switches became the standard fixed-configuration 100G platform in early cloud and enterprise data centers. They remain widely deployed in stable environments where 40G/100G leaf-spine fabrics continue to operate reliably.

Representative models: DCS-7060CX-32S-F and DCS-7060CX-32S-R.

4.2 Tomahawk+ (TH+) — 7060CX2 and 7060SX2

Tomahawk+ is an enhanced revision of the original Tomahawk, offering:

• Native IEEE 25GbE support (critical for the industry transition from 10G to 25G server NICs)
• Increased shared buffer from 16 MB to 22 MB
• Improved ECMP hashing and load balancing behavior
• Better power efficiency for dense 25G/100G deployments

TH+ powers the 7060CX2-32S (32 × 100G QSFP) and 7060SX2-48YC6 (48 × 25G SFP + 6 × 100G QSFP) platforms. These models marked the 7060 family’s transition into the 25GbE era and remain the most widely deployed 7060 generation in production today.

The CX2 and SX2 are ideal for organizations that have standardized on 25G server NICs but still run 100G uplinks to the spine layer. Their 450 ns latency and 22 MB DBA buffer provide predictable forwarding for general-purpose east-west traffic.

Representative models: DCS-7060CX2-32S-F and DCS-7060CX2-32S-R.

4.3 Tomahawk 3 (TH3) — 7060X4 (DX4 and PX4)

Tomahawk 3 (BCM56980) represents a major architectural leap, doubling the per-ASIC bandwidth to 12.8 Tbps and introducing native 400GbE port support. The 7060PX4-32 and 7060DX4-32 are single-ASIC platforms, delivering:

• 12.8 Tbps switching capacity (32 × 400G)
• 64 MB shared packet buffer — a 3× increase over TH+
• Cut-through latency from 700 ns for 400G
• Up to 8 billion packets per second
• 128-way ECMP for hyperscale fabric designs
• Advanced Congestion Management for NVMe and AI workloads

TH3 introduced the 7060 platform to the 400GbE era. The 7060PX4-32 uses OSFP interfaces while the 7060DX4-32 uses QSFP-DD — both deliver identical switching performance but accommodate different optics ecosystems (see Section 11 for OSFP vs QSFP-DD guidance).

Representative models: DCS-7060PX4-32-F, DCS-7060PX4-32-R, DCS-7060DX4-32-F, and DCS-7060DX4-32-R.

4.4 Tomahawk 4 (TH4) — 7060X5 (DX5, PX5, CX5)

Each TH4 ASIC (BCM56990) delivers 12.8 Tbps — the same per-chip bandwidth as TH3. However, the larger X5 models (DX5-64S, DX5-64E, DX5-64, PX5-64E) use a two-ASIC architecture to achieve 25.6 Tbps aggregate switching capacity in a single system. Smaller models like the DX5-32 and CX5-56D8 use a single TH4 ASIC at 12.8 Tbps or 8.8 Tbps respectively.

Key improvements over TH3:

• Up to 25.6 Tbps system throughput (dual-ASIC models) with native 800G port support
• Up to 114 MB shared packet buffer (dual-ASIC 2RU models) — nearly double TH3’s 64 MB
• Up to 10.6 billion packets per second
• Cut-through latency from 825 ns for 400G — higher than TH3’s 700 ns due to the increased pipeline complexity of the denser TH4 forwarding engine (see Section 10 for latency discussion)
• Up to 64 GB system memory (8× the TH3 generation) for large-scale state and telemetry
• 128-way ECMP with Dynamic Load Balancing (DLB) and Cluster Load Balancing (CLB)
• AI Analyzer for hardware-based traffic pattern analysis at 100-microsecond granularity

The X5 family is the broadest 7060 generation to date, spanning six models across three subfamilies (DX5, PX5, CX5). This range provides options from compact 1RU single-ASIC 12.8 Tbps switches to dense 2RU dual-ASIC 25.6 Tbps platforms with 800G-capable ports.

Representative models: DCS-7060DX5-64S-F, DCS-7060DX5-64E-F, DCS-7060DX5-32-F, DCS-7060PX5-64E-F, and DCS-7060CX5-56D8-F.

4.5 Tomahawk 5 (TH5) — 7060X6

Tomahawk 5 doubles the aggregate bandwidth again, enabling 51.2 Tbps switching in a single system and introducing purpose-built AI/ML networking features:

• 51.2 Tbps switching capacity — 64 × 800G OSFP ports in 2RU
• 165 MB shared packet buffer with Advanced Queueing support
• Up to 21.2 billion packets per second
• Cut-through latency from 700 ns
• Native RoCEv2 support for RDMA workloads
• Linear Pluggable Optics (LPO) support for lower power and improved reliability
• Packet Spraying for optimal load distribution across AI fabric links
• Enhanced congestion control: DCQCN, ECN, PFC with latency-based and throughput-based dynamic marking
• Field-replaceable supervisor card in 2RU variants — a first for the 7060 platform

The X6 generation represents the most significant architectural advance in the 7060 family’s history. With 51.2 Tbps in 2RU and a feature set specifically designed for GPU cluster fabrics, the 7060X6 positions the platform to serve as both leaf and spine in dedicated AI/ML networks.

Representative models: DCS-7060X6-64PE-F, DCS-7060X6-64PE-B-F, and DCS-7060X6-32PE-F.

4.6 Why Silicon Generation Matters for Long-Term Planning

When choosing 7060 platforms, match the silicon generation to your expected growth trajectory:

• Tomahawk / TH+ → Best for stable 10G/25G/100G environments not planning a speed migration in the near term
• TH3 → Standard for current 400G leaf-spine deployments with moderate AI/ML requirements
• TH4 → Ideal for 400G/800G fabrics requiring high buffer capacity, DLB, and future 800G readiness
• TH5 → Required for large-scale 800G AI/ML cluster interconnects with RoCEv2, packet spraying, and advanced congestion control

If you are unsure which silicon generation aligns with your requirements, BrightStar Systems can help evaluate workload profiles, growth plans, and optics constraints to select the ideal models.

5. 7060CX and 7060CX-32S (First-Generation Tomahawk, 40G/100G)

The original 7060CX models were among Arista’s first 100GbE-capable fixed-configuration switches. Built on first-generation Tomahawk silicon, the 7060CX-32S delivers 32 × 100G QSFP ports in a compact 1RU form factor with 6.4 Tbps of switching capacity.

Key characteristics of the CX platform:

• 32 × QSFP100 ports with full breakout flexibility (100G, 40G, 4×25G, 4×10G)
• 16 MB shared packet buffer with Dynamic Buffer Allocation
• 450 ns cut-through latency
• Up to 130 × 10GbE ports when fully broken out
• Reversible airflow (front-to-rear and rear-to-front SKUs available)
• 187W typical power draw — among the most power-efficient 100G platforms

The CX-32S remains relevant for expansions and replacements in established 40G/100G environments. Many organizations continue to operate large CX-32S fleets in stable leaf-spine fabrics where the platform’s low power consumption, proven reliability, and full EOS feature support make replacement unnecessary.

5.1 Best Use Cases for 7060CX-32S

• Legacy 40G/100G leaf-spine fabrics requiring expansion or like-for-like replacement
• Small-to-mid-size data centers running stable 10G/40G server populations
• Lab and staging environments where modern 400G bandwidth is not required
• Budget-conscious deployments where refurbished CX-32S units offer strong per-port economics

6. 7060CX2 and 7060SX2 (Tomahawk+, 25G/100G Leaf)

The 7060CX2 and 7060SX2 families represent the most widely deployed generation of the 7060 platform. These switches marked the industry’s transition from 10G to 25G server-facing connectivity and remain the standard leaf switch in a large number of production EVPN-VXLAN fabrics.

6.1 7060CX2-32S — 100G-Centric Leaf and Small Spine

The CX2-32S delivers the same 32 × QSFP100 port layout as the original CX-32S but with Tomahawk+ silicon improvements:

• 22 MB shared buffer (up from 16 MB) with Dynamic Buffer Allocation
• Native IEEE 25GbE breakout support (4×25G per QSFP)
• 128 × 25GbE maximum port count when fully broken out
• Same 450 ns cut-through latency as the CX-32S
• EOS Group 2 licensing

The CX2-32S is commonly deployed as a 100G leaf switch with 4×25G breakouts to servers, or as a compact spine in smaller leaf-spine fabrics.

6.2 7060SX2-48YC6 — Purpose-Built 25G Leaf

The SX2-48YC6 is designed explicitly for high-density 25G server access:

• 48 × 25G SFP ports for direct server connectivity (also support 10G and 1G)
• 6 × 100G QSFP uplink ports for spine connections
• 3.6 Tbps aggregate throughput
• 22 MB shared buffer with DBA
• 450 ns cut-through latency

The SX2-48YC6 eliminates the need for QSFP-to-SFP breakout cables, simplifying cabling and reducing per-port optics costs in 25G leaf deployments. This makes it the preferred choice when the primary requirement is dense, native SFP25 server access.

6.3 CX2/SX2 in EVPN Leaf-Spine Designs

                   +-----------------------------------------------+
                   |               100G Spine Layer                 |
                   |          7060CX2-32S or 7280CR Spine           |
                   +------------------+----------------------------+
                                      |
                              100G EVPN Underlay
                                      |
                 +--------------------+------------------------+
                 |                                             |
      +----------+-----------+                  +--------------+----------+
      | 7060CX2-32S-F        |                  | 7060SX2-48YC6           |
      | 100G Leaf VTEP       |                  | 25G Server Leaf VTEP    |
      | (4×25G breakouts)    |                  | (native SFP25 ports)    |
      +-----------+----------+                  +--------------+----------+
                  |                                            |
                25G                                          25G
                  |                                            |
           +------+------+                             +-------+-------+
           |  Compute    |                             |  Compute      |
           +-------------+                             +---------------+

In these designs, CX2 and SX2 switches act as EVPN VTEP leaf nodes with the spine providing underlay routing. Both platforms support symmetric IRB, VXLAN Type 2/3/5 routes, and MLAG for dual-homed servers.

6.4 Best Use Cases for CX2/SX2

• Standard 25G compute racks in enterprise and cloud environments
• General-purpose EVPN-VXLAN fabrics at moderate scale
• Mixed 10G/25G environments using SFP port flexibility
• Cost-effective leaf layer where 400G bandwidth is not yet required

7. 7060X4: DX4 and PX4 Families (Tomahawk 3, 400G)

The 7060X4 family extends the 7060 platform into the 400GbE era. Built on the Broadcom BCM56980 Tomahawk 3 ASIC, these switches deliver 12.8 Tbps of switching capacity per ASIC — providing the highest-density 400G connectivity available in a 1RU fixed-configuration switch.

The X4 family includes two models that differ only in their optics form factor:

• 7060PX4-32 — 32 × 400G OSFP ports + 2 × SFP+
• 7060DX4-32 — 32 × 400G QSFP-DD ports + 2 × SFP+

Both models share identical switching performance: 12.8 Tbps throughput from a single TH3 ASIC, 8 billion packets per second, 700 ns cut-through latency, and 64 MB of shared packet buffer. The choice between PX4 and DX4 is driven entirely by optics ecosystem alignment (see Section 11).

7.1 Key Capabilities

• 128-way ECMP with 64-way MLAG for hyperscale fabric designs
• Multi-speed flexibility: every 400G port supports breakout to 4×100G, 2×200G, 4×25G, or 4×10G
• Up to 128 × 100GbE interfaces when fully broken out
• Up to 640K IPv4 LPM routes and 300K IPv6 LPM routes (maximum values; actual capacity depends on EOS UFT profile mode)
• AI Analyzer for hardware-level ECMP traffic pattern analysis
• 1+1 redundant hot-swappable PSUs and N+1 hot-swappable fans
• Typical power draw under 17W per 400G port

7.2 X4 in 400G Leaf-Spine Designs

                   +-----------------------------------------------+
                   |               400G Spine Layer                 |
                   |       7060PX4-32 / 7060DX4-32 Spine            |
                   +------------------+----------------------------+
                                      |
                             400G EVPN Underlay
                                      |
                 +--------------------+------------------------+
                 |                                             |
      +----------+-----------+                  +--------------+----------+
      | 7060DX4-32-F         |                  | 7060PX4-32-F            |
      | 400G Leaf VTEP       |                  | 400G Leaf VTEP          |
      | (QSFP-DD)            |                  | (OSFP)                  |
      +-----------+----------+                  +--------------+----------+
                  |                                            |
         4×100G breakout                              4×100G breakout
                  |                                            |
           +------+------+                             +-------+-------+
           | GPU Nodes   |                             | Compute Racks |
           +-------------+                             +---------------+

7.3 Best Use Cases for 7060X4

• 400G leaf-spine fabrics for enterprise and cloud data centers
• AI/ML training cluster interconnects at moderate scale
• 100G-to-400G migration using breakout modes for investment protection
• High-bandwidth leaf in compute-dense environments

8. 7060X5: DX5, PX5, and CX5 Families (Tomahawk 4, 400G/800G)

The 7060X5 family is the broadest and most versatile generation in the 7060 lineup. Built on Tomahawk 4 silicon, the X5 spans six distinct models across three subfamilies, covering use cases from compact 1RU 12.8 Tbps switches to dense 2RU 25.6 Tbps platforms with 800G-capable ports.

8.1 Model Lineup

8.2 What Makes X5 Different from X4

While X4’s single TH3 ASIC delivers 12.8 Tbps, the dual-ASIC X5 models double that to 25.6 Tbps. Beyond raw throughput, TH4 silicon introduces several critical improvements:

• Native 800G port support (2×400G per port) for future bandwidth growth
• 114 MB buffer (vs 64 MB on X4) — significantly better burst absorption
• 64 GB system memory (vs 8 GB on X4) — enabling larger routing tables, more extensive telemetry, and richer on-box analytics
• Dynamic Load Balancing (DLB) — real-time load-aware flow placement that redistributes active and new flows to the least-loaded paths
• Cluster Load Balancing (CLB) — RDMA-aware flow placement that simultaneously optimizes leaf-to-spine and spine-to-leaf paths in AI clusters
• AI Analyzer — hardware-based ECMP utilization analysis at 100-microsecond granularity

8.3 The CX5-56D8: Bridging 100G and 400G

The DCS-7060CX5-56D8 occupies a unique niche in the X5 lineup. With 56 × 100G QSFP ports and 8 × 400G QSFP-DD uplinks, it provides a natural migration path for organizations running large 100G server-facing fabrics that need 400G spine connectivity. The CX5-56D8 allows operators to maintain existing 100G optics investments at the server edge while upgrading uplink bandwidth to 400G.

8.4 Best Use Cases for 7060X5

• High-density 400G leaf-spine fabrics with 64-port density per switch (DX5-64S)
• 800G-ready spine platforms that can start at 400G and grow to 800G without hardware replacement (DX5-64E, PX5-64E)
• AI/ML training clusters requiring DLB, CLB, and AI Analyzer for optimal RDMA performance
• 100G-to-400G migration using the CX5-56D8 as a transition leaf
• Compact 400G deployments using the 1RU DX5-32 for space-constrained environments

9. 7060X6 Family (Tomahawk 5, 800G AI/ML Platforms)

The 7060X6 family represents the current state of the art in the 7060 platform. Built on Tomahawk 5 silicon, these switches are designed from the ground up for 800GbE AI/ML cluster interconnects and next-generation hyperscale data center fabrics.

9.1 Model Lineup

• 7060X6-64PE — 64 × 800G OSFP + 2 SFP+ in 2RU; 51.2 Tbps; up to 320 interfaces
• 7060X6-64PE-B — 64 × 800G OSFP + 2 SFP+ in 2RU; 51.2 Tbps; up to 512 interfaces (enhanced breakout)
• 7060X6-32PE — 32 × 800G OSFP + 2 SFP+ in 1RU; 25.6 Tbps; up to 256 interfaces

9.2 AI/ML-Specific Architecture

The X6 introduces several features specifically engineered for GPU cluster interconnects:

Enhanced hardware-native RoCEv2 (RDMA over Converged Ethernet v2) — While RoCEv2 is supported across X4 and X5 platforms through EOS, the X6’s TH5 silicon adds enhanced hardware-native priority flow control and congestion management purpose-built for RoCEv2 at scale. This includes deeper hardware queue management and tighter integration with DCQCN marking — critical for maintaining lossless RDMA transport across large GPU clusters where thousands of queue pairs are active simultaneously.

Packet Spraying — Rather than using traditional per-flow ECMP hashing (which can create hotspots when flows are large and few), packet spraying distributes individual packets across all available paths. This is particularly effective in AI training workloads where collective operations generate massive, synchronized bursts.

Dynamic Load Balancing (DLB) and Cluster Load Balancing (CLB) — DLB monitors real-time link loads and dynamically reassigns flows to underutilized paths. CLB extends this with RDMA-aware placement that optimizes both leaf-to-spine and spine-to-leaf paths simultaneously — critical for reducing tail latency in distributed training jobs.

Advanced DCQCN — Enhanced Data Center Quantized Congestion Notification provides granular, hardware-driven congestion signaling with latency-based, throughput-based, and dynamic marking modes for scalable congestion control in large AI fabrics.

AI Analyzer — Hardware-based analysis of ECMP member utilization at intervals as fine as 100 microseconds, enabling operators to detect and correct load imbalances that software-based counters would miss entirely.

Linear Pluggable Optics (LPO) — Support for next-generation optics that consume less power and offer improved signal integrity compared to traditional pluggable modules, reducing per-port power consumption in large-scale 800G deployments.

9.3 Field-Replaceable Supervisor Card

A notable hardware innovation in the X6 is the field-replaceable supervisor card in the 2RU variants. The supervisor can be accessed by removing the two right-most fan modules, enabling control plane upgrades without replacing the entire switch. This is a first for the 7060 platform and significantly reduces long-term maintenance costs and downtime.

9.4 X6 in AI/ML Two-Tier Fabric Design

                   +-----------------------------------------------+
                   |             800G Spine Layer (32 Spines)       |
                   |              7060X6-64PE (64 × 800G each)      |
                   +------------------+----------------------------+
                                      |
                             800G EVPN Underlay
                                      |
     +--------------------------------+-------------------------------+
     |                |                |                |              |
+----+----+    +------+----+    +------+----+    +-----+-----+  ... ×128
| X6 Leaf |    | X6 Leaf   |    | X6 Leaf   |    | X6 Leaf   |
| 64×800G |    | 64×800G   |    | 64×800G   |    | 64×800G   |
| (total) |    | (total)   |    | (total)   |    | (total)   |
+---------+    +-----------+    +-----------+    +-----------+
     |                |                |                |
  32 ports          32 ports         32 ports         32 ports
  to spine          to spine         to spine         to spine
  32 ports          32 ports         32 ports         32 ports
  to xPU/GPU        to xPU/GPU       to xPU/GPU       to xPU/GPU

              4,096 xPU Host Ports at 800G Each

In this architecture, 7060X6-64PE switches serve as both leaf and spine nodes in a dedicated two-tier AI fabric. Each X6-64PE leaf has 64 × 800G ports total — in a typical 1:1 oversubscription design, 32 ports connect to the spine layer and 32 ports connect to GPU/xPU endpoints at 800G each. With 128 leaf switches, the fabric provides 4,096 host-facing 800G ports. For higher host density per leaf, alternative oversubscription ratios (e.g., 48 host + 16 spine) can be configured based on traffic profile and acceptable oversubscription level.

9.5 Best Use Cases for 7060X6

• Large-scale AI/ML training clusters requiring 800G per-GPU connectivity
• Next-generation hyperscale spine for 400G/800G leaf-spine fabrics
• High-performance computing (HPC) with RDMA/RoCEv2 requirements
• Dense 800G aggregation in cloud and colocation environments

10. Latency, Buffering, and Dynamic Buffer Allocation

A defining characteristic of the 7060 platform is its emphasis on low, deterministic latency combined with dynamically allocated packet buffers. Understanding the tradeoffs between latency and buffer depth is essential for matching the right 7060 model to your traffic profile.

10.1 Latency Across Generations

Understanding the X5 latency regression: The 7060X5 is the only 7060 generation where cut-through latency increased versus the prior generation. TH4’s 825 ns (compared to TH3’s 700 ns) is a direct consequence of the more complex forwarding pipeline required to support higher port density and the inter-ASIC fabric in dual-ASIC models. In practice, this 125 ns delta is negligible for most data center workloads — it is well within the sub-microsecond envelope that AI/ML and east-west traffic patterns require. The TH5 (X6) resolved this by returning to 700 ns while simultaneously doubling throughput to 51.2 Tbps.

10.2 Dynamic Buffer Allocation (DBA)

All 7060 platforms use Dynamic Buffer Allocation rather than fixed per-port buffers. In cut-through mode, packets traverse the switch pipeline at the latency figures shown above. When congestion occurs on a specific egress port, the shared buffer pool dynamically allocates additional memory to that port to absorb the burst.

This approach is fundamentally different from the fixed per-port buffer architectures used in some competing platforms. DBA ensures that idle ports do not waste buffer resources, while congested ports can temporarily access a much larger portion of the total buffer pool — providing significantly better burst absorption than the raw buffer number might suggest.

10.3 When Shallow Buffers Are and Are Not Appropriate

The 7060’s buffer profile (16–165 MB depending on generation) is designed for environments where traffic patterns are relatively well understood and east-west flows dominate. This includes most leaf-spine fabrics, compute clusters, and AI/ML interconnects where traffic patterns are synchronized and predictable.

For environments with heavy fan-in congestion, unpredictable microburst patterns, or large-scale storage traffic (iSCSI, NFS, NVMe/TCP), deeper-buffer platforms like the Arista 7280R series may be more appropriate at aggregation or spine layers. Many production networks combine 7060 leaf switches with 7280R spine/aggregation switches to balance latency and burst absorption.

11. OSFP vs QSFP-DD: Choosing the Right Optics Form Factor

Starting with the X4 generation, the 7060 family offers models with both OSFP and QSFP-DD front-panel interfaces. Both form factors support the same line rates (400G and 800G), but they differ in physical design and thermal characteristics.

QSFP-DD (Quad Small Form Factor Pluggable — Double Density)

• Backward compatible with existing QSFP28 and QSFP56 optics
• Widely supported across multi-vendor environments
• Preferred when existing 100G/200G QSFP optics inventory needs to be reused
• 7060 models: DX4-32, DX5-32, DX5-64, DX5-64S, DX5-64E, CX5-56D8

OSFP (Octal Small Form Factor Pluggable)

• Larger form factor with better thermal dissipation for high-power optics
• Supports Linear Pluggable Optics (LPO) on X6 platforms
• Preferred for 800G deployments where optical power budgets are higher
• 7060 models: PX4-32, PX5-64E, X6-64PE, X6-64PE-B, X6-32PE

Practical selection guidance: If your environment has an existing QSFP optics inventory or requires backward compatibility with 100G QSFP28 modules, choose DX models. If you are building a new 400G/800G fabric where thermal headroom and LPO support are priorities, choose PX/X6 models with OSFP.

12. Common Deployment Architectures

The 7060 series is most commonly deployed in three primary roles within data center networks:

12.1 Leaf / Top-of-Rack (ToR)

This is the most common deployment role for the 7060. As a leaf switch, the 7060 provides high-density server-facing ports with uplinks to a spine layer. The specific model depends on the server interface speed:

• 25G servers → 7060SX2-48YC6 or 7060CX2-32S (with breakouts)
• 100G servers → 7060CX5-56D8 or 7060DX5-64S (with breakouts)
• 400G endpoints → 7060DX4-32, 7060DX5-32, or 7060DX5-64S
• 800G endpoints (GPU/xPU) → 7060X6-64PE or 7060DX5-64E / PX5-64E

12.2 Fixed-Configuration Spine

In small-to-mid-size fabrics, 7060 switches can also serve as spine nodes. This is common in deployments where the total fabric size does not require modular chassis platforms:

• 100G spine → 7060CX2-32S (32 × 100G)
• 400G spine → 7060PX4-32 or 7060DX4-32 (32 × 400G)
• 800G spine → 7060X6-64PE (64 × 800G) or 7060X6-32PE (32 × 800G)

12.3 AI/ML Cluster Fabric

The X5 and X6 generations are increasingly deployed as dedicated AI/ML cluster interconnects, where the 7060 serves as both leaf and spine in a purpose-built two-tier fabric for GPU or xPU clusters (see Section 14 for detailed AI fabric designs).

13. EVPN-VXLAN Fabric Designs with 7060 Platforms

All 7060 platforms support EVPN-VXLAN fabric designs through Arista EOS, enabling modern overlay network architectures with distributed anycast gateway routing, multi-tenancy, and workload mobility.

13.1 EVPN Route Types Supported

7060 switches handle all standard EVPN route types:

• Type 2 — MAC/IP Advertisement for host learning and mobility
• Type 3 — Inclusive Multicast Ethernet Tag for BUM traffic handling
• Type 5 — IP Prefix Route for distributed inter-subnet routing

13.2 Symmetric vs Asymmetric IRB

Arista EOS supports both IRB models on 7060 platforms:

• Asymmetric IRB — Simpler to configure but less scalable; routing is performed only at the ingress leaf
• Symmetric IRB — More scalable; uses a transit VNI so that both ingress and egress leaf nodes participate in routing

For most production deployments, symmetric IRB is recommended, especially when deploying multiple VRFs or scaling beyond a small number of racks.

13.3 MLAG and Dual-Homing

All 7060 platforms support Multi-Chassis Link Aggregation (MLAG) for dual-homed server connectivity. MLAG allows a pair of 7060 leaf switches to present a single logical LAG to a downstream server or appliance, providing link-level redundancy without requiring the server to run a routing protocol.

Key MLAG specifications across the 7060 family:

• Up to 64 ports per MLAG
• Up to 128 LAG groups per system
• Active/active forwarding on both MLAG peers

13.4 Practical EVPN Fabric Design with 7060

                          +-------------------------------------------+
                          |              Spine Layer                  |
                          |      7060DX4-32 or 7280CR3 Spine          |
                          |      (EVPN Route Reflector optional)      |
                          +----------+--------------+-----------+----+
                                     |              |           |
                             400G EVPN Underlay (eBGP or iBGP)
                                     |              |           |
              +----------------------+    +---------+    +------+-----------+
              |                           |              |                  |
   +----------+----------+    +-----------+-+    +-------+--------+   +----+-----------+
   | 7060DX5-64S-F       |    | 7060DX4-32-F |    | 7060CX5-56D8-F |   | 7060SX2-48YC6 |
   | VTEP: 400G Leaf     |    | VTEP: 400G   |    | VTEP: 100G+400G|   | VTEP: 25G Leaf|
   +----------+----------+    +------+-------+    +--------+-------+   +-------+-------+
              |                      |                     |                   |
           400G                   4×100G                 100G                 25G
              |                      |                     |                   |
        +-----+-----+        +------+------+        +-----+-----+      +-----+-----+
        | GPU Cluster|        | Compute     |        | Compute   |      | Compute   |
        +------------+        +-------------+        +-----------+      +-----------+

This design illustrates a mixed-generation EVPN fabric where different 7060 models serve as leaf VTEPs based on the server interface speed requirements of each rack or pod. EOS provides a consistent control plane across all leaf generations.

14. AI/ML and HPC Fabric Design

The 7060X5 and X6 generations have introduced features specifically designed for the unique traffic patterns of AI/ML training workloads. Understanding these patterns is essential for selecting the right platform and designing an effective AI fabric.

14.1 Why AI Training Traffic Is Different

AI training workloads — particularly distributed training using frameworks like PyTorch, TensorFlow, and JAX — generate traffic patterns that are fundamentally different from traditional data center traffic:

• Synchronized collective operations (AllReduce, AllGather, ReduceScatter) create massive, coordinated bursts across hundreds or thousands of GPU nodes simultaneously
• Few, large flows rather than many small flows — traditional per-flow ECMP hashing can create severe hotspots
• RDMA (RoCEv2) transport requires lossless Ethernet with precise congestion control
• Tail latency sensitivity — a single straggler GPU can delay the entire training step, making network-level tail latency a direct contributor to job completion time (JCT)

14.2 7060 Features for AI Fabrics

The following features, available on X5 and X6 platforms, directly address AI training requirements:

Packet Spraying (X6) distributes packets across all available ECMP paths rather than pinning entire flows to a single path. This eliminates the hotspot problem inherent in hash-based ECMP when flows are large and few.

Cluster Load Balancing (CLB) (X5, X6) uses RDMA queue pair awareness to optimize flow placement. Unlike traditional DLB which only optimizes leaf-to-spine paths, CLB simultaneously optimizes both leaf-to-spine and spine-to-leaf directions — critical because every spine typically has only one path to each destination leaf.

Dynamic Load Balancing (DLB) (X5, X6) monitors per-link utilization in real time and dynamically reassigns both new and existing flows to the least-loaded paths, reducing the probability of drops.

AI Analyzer (X4, X5, X6) — available across all TH3+ generations — collects ECMP member utilization data at hardware speed with granularity as fine as 100 microseconds. This enables operators to detect and diagnose load imbalances that software-based monitoring (typically polling at 1-second intervals) would entirely miss.

Enhanced DCQCN (X6) provides advanced congestion notification with latency-based, throughput-based, and dynamic marking modes, enabling fine-grained congestion control tuned specifically for RDMA traffic profiles.

14.3 Two-Tier AI Fabric Reference Design

                              Spine: 32 × 7060X6-64PE
                    (each spine: 64 × 800G OSFP, 51.2 Tbps)
                   +----+----+----+----+----+----+----+----+
                   | S1 | S2 | S3 | S4 | .. | .. | .. |S32 |
                   +--+-+--+-+--+-+--+-+--+-+--+-+--+-+--+-+
                      |    |    |    |    |    |    |    |
                      +----+----+----+----+----+----+----+
                      |    Full-mesh 800G interconnect   |
                      +----+----+----+----+----+----+----+
                      |    |    |    |    |    |    |    |
                   +--+-+--+-+--+-+--+-+--+-+--+-+ ...  +
                   | L1 | L2 | L3 | L4 | .. | .. |     |L128|
                   +--+-+--+-+--+-+--+-+--+-+--+-+     +--+-+
                      |    |    |    |    |    |           |
                    Leaf: 128 × 7060X6-64PE
                   (each leaf: 64 × 800G — 32 to spine, 32 to xPU)

                   32 GPU/xPU per leaf × 128 leaves
                   = 4,096 GPU/xPU endpoints at 800G each

This two-tier design uses 7060X6-64PE switches at both leaf and spine. Each leaf has 64 × 800G ports: in a balanced 1:1 ratio, 32 ports connect to the 32 spine switches and 32 ports connect to GPU/xPU endpoints at 800G each. With 128 leaf switches, the fabric provides 4,096 host-facing 800G ports with non-blocking bandwidth between any pair of nodes. Higher host-to-spine oversubscription ratios (e.g., 48 host ports + 16 spine uplinks at 3:1) can extend endpoint count per leaf at the cost of bisection bandwidth.

15. Storage, Microbursts, and East-West Traffic

The 7060 platform is optimized for east-west traffic — the dominant pattern in modern spine-leaf data centers where most traffic flows laterally between servers rather than vertically to external networks.

15.1 When the 7060 Is the Right Choice for Storage

For storage workloads with moderate burst profiles and predictable traffic patterns, the 7060’s DBA buffer architecture provides effective burst absorption:

• NVMe/TCP workloads with well-characterized flow patterns
• Distributed storage systems (Ceph, vSAN, MinIO) in compute-adjacent leaf roles
• AI/ML data pipeline ingestion where throughput consistency matters more than deep burst absorption

15.2 When Deeper Buffers Are Needed

For storage environments with heavy fan-in patterns, large-scale iSCSI, or NFS workloads with unpredictable burst characteristics, the 7280R series provides deep VOQ buffers (hundreds of megabytes) that absorb sustained congestion without packet loss.

A common design pattern combines 7060 leaf switches at the ToR (handling server-facing connectivity) with 7280R spine or aggregation switches (providing deep-buffer protection at congestion points). This approach gives each layer the characteristics best suited to its traffic profile.

16. Migration Paths: 10G → 25G → 100G → 400G → 800G

The 7060 platform’s generational consistency makes it uniquely well suited for phased speed migrations. Because every generation runs the same Arista EOS with the same operational model, organizations can upgrade hardware incrementally without retraining operations teams or rebuilding automation tooling.

16.1 Common Migration Sequences

10G → 25G (Server NIC upgrade)
Replace 7060CX-32S or SX2 leaf switches with CX2-32S units. Servers upgrade from 10G SFP+ to 25G SFP28 NICs. Existing 100G spine infrastructure remains unchanged.

25G/100G → 400G (Fabric bandwidth upgrade)
Deploy 7060DX4-32 or PX4-32 as new spine nodes, then gradually migrate leaf switches from CX2/SX2 to X4 or X5 models. Use QSFP-DD breakout (4×100G) on new switches to maintain connectivity to existing 100G infrastructure during transition.

400G → 800G (AI/ML scale-out)
Deploy 7060X5 or X6 platforms as 800G-ready leaf and spine nodes. Start with 400G optics in 800G-capable ports, then upgrade to 800G optics as GPU node density increases. The DX5-64E, PX5-64E, and X6 families all support this 400G-to-800G optics migration without hardware replacement.

16.2 Breakout Modes for Mixed-Speed Environments

Every 7060 generation supports extensive port breakout configurations that enable mixed-speed environments during migration:

• 400G QSFP-DD → 4×100G, 2×200G, 4×25G, 4×10G
• 800G QSFP-DD → 2×400G, 4×200G, 8×100G
• 100G QSFP → 4×25G, 4×10G, 2×50G

These breakout modes allow a single switch to simultaneously serve endpoints at different speeds, eliminating the need for one-to-one hardware-to-speed matching during transitional periods.

17. Model Selection Framework

Selecting the right 7060 model requires matching several key variables to your deployment requirements:

1. What is the server/endpoint interface speed?

• 10G/25G → CX2-32S or SX2-48YC6
• 100G → CX5-56D8 (native 100G ports) or DX5-64S (100G breakout from 400G)
• 400G → DX4-32, PX4-32, DX5-32, or DX5-64S
• 800G → DX5-64E, PX5-64E, X6-64PE, or X6-32PE

2. What is the uplink/spine interface speed?

• 100G uplinks → CX2-32S, SX2-48YC6
• 400G uplinks → DX4-32, PX4-32, DX5-32, DX5-64S, CX5-56D8
• 800G uplinks → DX5-64E, PX5-64E, X6-64PE, X6-32PE

3. Do you need AI/ML-specific features (DLB, CLB, RoCEv2, packet spraying)?

• Basic AI/ML → X4 (AI Analyzer) or X5 (adds DLB, CLB)
• Advanced AI/ML → X6 (adds enhanced hardware-native RoCEv2, packet spraying, DCQCN, LPO)

4. What is the form factor constraint?

• 1RU → CX-32S, CX2-32S, SX2-48YC6, DX4-32, PX4-32, DX5-32, DX5-64E, PX5-64E, X6-32PE
• 2RU → DX5-64S, DX5-64, CX5-56D8, X6-64PE, X6-64PE-B

5. OSFP or QSFP-DD?

• Existing QSFP optics inventory → DX models
• New build with 800G / LPO priority → PX or X6 models

18. Generational Comparison: Key Specs Across 7060 Families

*IPv4 LPM route capacity depends on the EOS Unified Forwarding Table (UFT) profile mode configured on the switch. The figures above represent maximum values achievable in LPM-optimized profiles. Allocating more table space to LPM reduces capacity available for other forwarding resources (ARP/ND, ACLs, etc.). Consult the Arista EOS configuration guide for your specific platform and EOS release to determine the profile that best fits your deployment.
**Max Speed reflects the native per-port line rate. The DX5-64S QSFP-DD ports are natively 400G but support 2×400G breakout configurations that can connect to 800G-capable endpoints via port channeling. For deployments planning future 800G connectivity from QSFP-DD platforms, this breakout capability may be a relevant consideration.

19. 7060 Series vs Arista 7280 Series: Architectural Differences

The 7060 and 7280 are often deployed together in the same fabric, but they are architecturally distinct platforms designed for different roles. Understanding these differences prevents misapplication and ensures each layer of the fabric gets the characteristics it needs.

The bottom line: Use the 7060 where you need low latency, high port density, and predictable east-west forwarding. Use the 7280R where you need deep buffering, high routing scale, and congestion resilience. In many fabrics, the optimal design uses both — 7060 at the leaf and 7280R at the spine or aggregation layer.

For a comprehensive guide to the 7280R platform, see our Arista 7280R Series architecture guide.

20. Buying Used and Refurbished Arista 7060 Switches

The 7060 platform’s proven reliability and long EOS support lifecycle make it well suited for the secondary market. Many organizations successfully deploy refurbished 7060 switches in production environments, achieving significant cost savings without compromising on performance or supportability.

20.1 What to Evaluate When Buying Refurbished

• Hardware revision — Confirm the unit’s hardware revision supports the EOS version required for your features. Some early revisions of CX/CX2 models may not support the latest EOS releases.
• Airflow direction — Verify that the -F (front-to-rear) or -R (rear-to-front) airflow matches your rack cooling design. Mixing airflow directions in the same rack creates hot spots and thermal issues.
• Power supply compatibility — Confirm AC vs DC power supply type and wattage for your facility.
• Optics compatibility — Ensure your existing optics inventory (QSFP28, QSFP-DD, OSFP, SFP+, SFP28) is compatible with the specific model’s port type.
• EOS license group — 7060 models span EOS license Groups 2, 3, and 4. Verify that the license group matches your feature requirements (EVPN, VXLAN, MPLS, etc.).

20.2 Why Refurbished 7060 Switches Make Sense

• No performance difference — A refurbished 7060CX2-32S runs the same EOS and delivers the same forwarding performance as a new unit
• Proven reliability — The 7060 platform has accumulated years of field-proven operation across thousands of deployments
• Excellent for lab, staging, and expansion — Refurbished units enable organizations to build out non-production environments or expand existing fabrics at a fraction of new-unit cost
• Sustainability — Extending the service life of existing hardware reduces electronic waste

21. Why Engineers Source 7060 Hardware from BrightStar Systems

BrightStar Systems specializes in sourcing, testing, and supplying enterprise networking hardware with a particular depth of expertise in Arista platforms. Our team works directly with data center engineers to ensure exact SKU matching, airflow correctness, optics compatibility, and EOS license alignment.

• Deep Arista expertise — We maintain detailed knowledge of every 7060 subfamily, revision, and SKU variant
• Exact SKU matching — We understand the difference between a DX4-32-F and a PX4-32-R and can help you source exactly what your fabric requires
• Tested and verified — Every unit is tested and verified before shipment
• BrightStar warranty — All hardware ships with our standard warranty coverage
• Technical consultation — We can advise on model selection, migration planning, and fabric design to help you get the most from your investment

22. Related Resources

• Used & Refurbished Arista 7060 Series Switches — Full Inventory
• Arista 7280R Series Architecture & Deployment Guide