As artificial intelligence (AI), machine learning (ML), and high-performance computing (HPC) continue to accelerate, modern data centers face unprecedented demands on bandwidth, latency, and scalability.
To keep pace with the explosive growth of GPU clusters and AI workloads, networks need switching platforms built for raw performance, deterministic latency, and seamless scale-out.
Today, we’re proud to introduce Creox’s next-generation AI Ethernet switches:
CS84-128Q1A (400G) and CS92-64OPA (800G).
Both platforms are powered by Broadcom’s BCM78900 Tomahawk 5 — the industry’s leading 51.2T switch silicon — delivering massive switching capacity, exceptional power efficiency, and the flexibility required for hyperscale AI networking.
Purpose-Built for AI and Hyperscale Data Centers
The CS84-128Q1A and CS92-64OPA are engineered for spine and leaf roles in massive AI and HPC topologies, where network performance directly impacts training speed and cluster efficiency.
Key capabilities include:
- Non-blocking Ethernet switching for predictable throughput in large GPU clusters
- High reliability with redundant cooling and power
- Full Layer 2/3 feature support for diverse deployment architectures
- Open networking flexibility, with ONIE pre-installed and support for SONiC and multiple commercial NOS options
This design approach gives operators the freedom to choose the network software stack that best fits their environment.
CS84-128Q1A — High-Density 400G AI Switch (128 × QSFP112)
The CS84-128Q1A is a 4RU ultra-high-bandwidth switch delivering 102.4Tbps full-duplex capacity across 128 QSFP112 ports.
It is purpose-built for dense leaf or spine layers in AI clusters, where predictable latency, reliable throughput, and efficient power delivery are essential.
Highlights
- 128 × 400GbE QSFP112 (supports 400G/200G/100G breakout)
- 102.4Tbps non-blocking switching capacity
- Redundant power: 2+2 CRPS, 2000W
- Redundant cooling: N+1 fan trays
- Intel Ice Lake-D 8-core CPU, 32GB DDR4 ECC, 240GB SSD
- Front-to-back airflow (F2B)
- Pre-loaded ONIE, supports SONiC and major commercial NOS
Ideal for:
AI cluster leaf/spine, large-scale HPC fabrics, hyperscale AI deployments
CS92-64OPA — Ultra-High Bandwidth 800G AI Switch (64 × OSFP112)
Designed for next-generation GPU cluster backbones, the CS92-64OPA delivers 64 × 800GbE ports in a space-efficient 2RU chassis — while maintaining the same 102.4Tbps of switching throughput.
Featuring a PHY-less design, it offers exceptional signal integrity, reduced latency, and minimal power overhead, making it ideal for large-scale AI training fabric and HPC environments.
Highlights
- 64 × 800GbE OSFP112 (8×112G PAM4) + 2 × SFP28
- 102.4Tbps non-blocking throughput
- PHY-less architecture for lower power and cleaner signal
- Intel Ice Lake-D 8-core CPU, 32GB DDR4 ECC, 480GB SSD
- 1+1 CRPS 3200W PSU, N+1 fan design
- 2RU form factor, air-cooled (F2B airflow)
- ONIE pre-installed, supports SONiC and commercial NOS
Ideal for:
GPU backbone switching, HPC compute clusters, 800G AI network fabrics
Technical Summary
Technical Summary
| Model | CS84-128Q1A | CS92-64OPA |
| Switch ASIC | Broadcom BCM78900 (Tomahawk 5) | Broadcom BCM78900 (Tomahawk 5) |
| Port Configuration | 128 × QSFP112 (4×112G PAM4) | 64 × OSFP112 (8×112G PAM4) + 2 × SFP28 |
| Switching Capacity | 102.4 Tbps (Full Duplex) | 102.4 Tbps (Full Duplex) |
| CPU | Intel Ice Lake-D 8C16T (D-1734NT) | Intel Ice Lake-D 8C16T (D-1734NT) |
| Memory / Storage | 2 × DDR4 16GB (up to 32GB) with ECC / 240GB M.2 SSD | 2 × DDR4 16GB (up to 32GB) with ECC / 480GB M.2 SSD |
| BMC | Aspeed AST2520 | Aspeed AST2620 |
| Cooling | N+1 Fans, Front-to-Back airflow | N+1 Fans, Front-to-Back airflow |
| Power Supply | 2+2 CRPS 2000W (AC/HVDC) | 1+1 CRPS 3200W (AC/HVDC) |
| Chassis Size | 4RU (440 × 800 × 175 mm) | 2RU (440 × 780 × 87.5 mm) |
| Weight | ~45 kg | ~30 kg |
| Operating Temp. | 0°C to 40°C | 0°C to 40°C |
| OS Support | ONIE, SONiC, Commercial NOS | ONIE, SONiC, Commercial NOS |
Empowering the Next Wave of AI Infrastructure
With Broadcom Tomahawk 5 technology at the core, the CS84-128Q1A and CS92-64OPA deliver the bandwidth, efficiency, and scalability required for next-generation AI and HPC architectures.
As GPU clusters grow larger and more interconnected, these switches provide:
- The ultra-high throughput needed for massive distributed training
- The deterministic low-latency fabric that AI workloads demand
- The flexibility of open networking to integrate seamlessly into modern data center stacks
Creox’s new 400G and 800G AI switches set the foundation for the future of hyperscale AI computing — enabling faster insights, greater efficiency, and unprecedented scale.
Unlock the full potential of your AI workloads.
Whether you’re scaling GPU clusters, modernizing your data center, or upgrading to 800G networking, Creox can help you get there faster.
Contact us now for pricing, lead times, and deployment support.
