Skip to main content

64×800G AI Switch FAQ: Performance, Compatibility, and Use in AI/ML/HPC

written by Asterfuison

July 11, 2025

Introduction

The CX864E-N is Asterfusion’s next-generation 64×800G AI switch powered by the Marvell Teralynx 10 chip, delivering 51.2Tbps of non-blocking throughput for AI/ML, cloud-scale, and HPC networks. In this FAQ, we answer the most frequently asked questions around architecture, latency, RoCEv2 support, optics compatibility, SONiC features, and TCO — everything decision-makers and network engineers need to know.

64x800G OSFP AI switch panel-1
64x800G OSFP AI switch panel-2

800G OSFP AI Switch Video


1. Performance & Architecture

Q: Is the 64×800G configuration based on OSFP or QSFP-DD800, and does it support backward compatibility with 400G?

A: CX864E-N AI switch is an OSFP form factor, 8×112G lanes backward compatible to 8×56G (i.e., 400G OSFP).

Q: What breakout configurations are supported? 2×400G / 8×100G / 4×200G?

A: All listed breakout modes are supported, but the peer device must support 112G SerDes.

Q: Is the AI switch 51.2T full-duplex and non-blocking? Is it a Clos architecture?

A: Yes. The CX864E-N is full-duplex and non-blocking. By single-chip pizza box design, Spine-Leaf architecture is more effective than chassis switches.

Q: How is forwarding performance under mixed traffic (large + small packets)?

A: Mixed traffic will not affect forwarding performance.

Q: The minimum latency is listed as 560ns — under what conditions? Is it cut-through mode?

A: The latency is measured under line-rate forwarding of standard Ethernet packets. Larger packets result in lower latency (around 500ns average for 9216-byte frames). The measurement is based on the store-and-forward mode, which means the cut-through performance can be even better.


2. AI/ML, HPC & Financial Scenario Suitability

Q: Has it passed NVIDIA RoCEv2 compatibility tests?

A: Yes. The 800G AI switch has passed the RoCEv2 compatibility test. The available report is tested with H20.

Q: Does CX864E-N support fine-grained ECN and PFC management? How many Priority Groups are supported?

A: Yes, it supports fine-grained ECN and PFC. Lossless queues have 2 Priority Groups.

Q: Does it support Inband Telemetry / INT? Can it be used for traffic visualization and bottleneck analysis?

A: Yes. CX864E-N AI switch supports INT. It can report metrics of BDC (Buffer Drop Collector) and HDC (High Delay Collector) to Prometheus, and bottleneck analysis can be realized via Grafana.

Q: Are metrics like Token Generation Rate (TGR) and Inter-Token Latency (ITL) available?

A: Yes. We have inference test reports. CX864E-N delivering 20% higher throughput (TGR) and significantly lower 90th percentile inference latency (P90 ITL). Refer to this: RoCE Beats InfiniBand: Asterfusion 800G Switch Crushes It with 20% Faster TGR. Please contact us for full report.

Q: In AI clusters, does the switch support QoS prioritization under concurrent jobs? Bandwidth guarantees?

A: Yes, the switch supports QoS prioritization and bandwidth guarantees in AI job concurrency scenarios.

Q: For financial/stock exchange applications, does it support IEEE 1588v2 (PTP) and Transparent Clock (TC)?

A: Yes. CX864E-N AI switch supports PTP/1588v2, including Boundary Clock (BC) and Transparent Clock (TC) roles, making it suitable for finance.


3. Optics, Power, and Thermal Design

Q: What 800G modules are supported?

We have independently developed and manufactured a range of 800G modules, as well as AOC and DAC cables, including QSFP-DD form factors. The product lineup includes 2VR4, VR8, 2FR4, DR8, and 2LR4 — all offered at competitive prices and with high performance. Please visit this page to explore: https://cloudswit.ch/product-category/optical-transceivers/

Q: Is PAM4 hardware FEC supported? Which FEC algorithm is used?

A: Yes, it supports PAM4 hardware FEC; it uses Reed-Solomon (RS).

Q: What are typical and maximum power consumption values? What fan/airflow solution is used?

A: Typical power consumption of CX864E-N is 2000W, and maximum power consumption is 2200W. It has 3+1 hot-swappable fans, uses front-to-rear airflow.

Q: How many 800G ZR modules can run simultaneously? Any thermal or downclocking protection?

A: Up to 32 ZR modules are recommended. The CX864E-N switch supports thermal protection.


4. SONiC System and Operability

Q: What custom optimizations are included in Asterfusion Enterprise SONiC for the 51.2T platform?

A: Optimizations include:

  • Adaptive Routing and Switching (ARS
  • Q-pair-based RDMA Hash
  • Packet spraying for better load balancing
  • These optimizations enable it to perform better in AI network scenarios.

Q: Is FRR + BGP-EVPN supported for VXLAN Fabric? Large-scale ECMP?

A: Yes. It supports BGP-EVPN, up to 4K VXLAN groups, and 256-way ECMP.

Q: Does it support gNMI, REST API, OpenConfig, etc.?

A: Yes, the CX-N series includes 800G switch supports standard management interface protocols such as gNMI, REST API, and NETCONF.

Q: Does the system support ISSU (In-Service Software Upgrade), NSR, or GR?

A: ISSU is not supported since it requires a dual control plane module in hardware. AsterNOS supports hot-patch upgrades via containers to minimize service interruption.


5. TCO & Commercial Considerations

Q: How does it compare to the Broadcom Tomahawk 5? What are Asterfusion’s advantages?

  • Lower latency: 560ns (Asterfusion) vs ~800ns (Broadcom Tomahawk 5)
  • Bigger buffer: 200MB (Asterfusion) vs 165MB (Broadcom Tomahawk 5)
  • These features make it particularly well-suited for today’s AI training, AI inference, and other related network scenarios.

Q: Do you offer integrated hardware-software bundles or OEM/ODM customization?

A: Yes. Asterfusion offers turnkey delivery of integrated software-hardware, and OEM customization is available (with MoQ).

Q: Is modular maintenance supported? Are fans and PSUs hot-swappable? Any redundancy?

A: Yes. The CX864E-N switch has 3+1 fans and 1+1 power supplies, all hot-swappable with redundancy.

Q: What’s the lead time? Can PoC units and remote test support be provided?

A: Lead time is 1 week. We will offer heavily discounted demo units to facilitate your evaluation. This can help accelerate the customer decision-making process, and you’re welcome to purchase these high-discount samples.


Conclusion

The CX864E-N 64×800G switch delivers the bandwidth, latency, and telemetry visibility required for AI/ML, HPC, and ultra-low latency environments. With robust SONiC support, RoCEv2 compatibility, and modular design, it’s an ideal building block for next-gen data centers.

👉 Still have questions? Contact our team or request a demo unit today.

Latest Posts