UfiSpace Unveils Full-Spectrum 1.6T AI Networking Portfolio Powered by Broadcom Tomahawk 6, Tomahawk 5 Ultra, and Qumran4D for Next-Gen AI Data Centers

Demonstrating five disaggregated open networking platforms engineered to optimize Total Cost of Ownership (TCO), lower PUE, and accelerate Scale-Up, Scale-Out and Scale-Across AI workloads.

TAIPEI CITY, Taiwan – August 11, 2026 - As global artificial intelligence (AI) models expand from 175 billion to over 1.5 trillion parameters, traditional data center network architectures face structural challenges regarding throughput, latency, and power management. To address these operational constraints, UfiSpace, a provider of open networking solutions, has introduced five high-performance open networking platforms spanning from 25.6 Tbps to 102.4 Tbps capacity under the showcase theme " Tomahawk 6 Liquid Cooling and Tomahawk 5 Ultra for Data Centers" at OCP APAC 2026.

Designed for cloud service providers, UfiSpace's expanded portfolio delivers field-proven disaggregated solutions tailored for three core AI networking dimensions: high-density AI Scale-Out training fabrics, tightly coupled Scale-Up GPU clusters, and geo-distributed Scale-Across data center interconnects (DCI).

Thermal Management at Scale: 102.4T Liquid-Cooled Switching
Modern high-density AI server racks now require cooling capacities reaching 130 kW to 140 kW per rack. Industry data from TrendForce projects that liquid cooling adoption will reach approximately 40% by 2026 as facilities transition toward Liquid-to-Liquid (L2L) architectures to optimize Power Usage Effectiveness (PUE). 

To solve thermal management bottlenecks at the high-radix network layer, UfiSpace presents the S9331-64HO3L, a 64-port 1.6T open data center switch built on Broadcom Tomahawk 6 silicon. Engineered to Open Rack v3 (ORv3) specifications in a 2OU form factor, the platform delivers 102.4 Tbps of switching capacity and supports direct liquid cooling to maintain thermal stability under maximum workload density. For air-cooled environments, UfiSpace offers the S9331-64HO in a 3RU chassis.

Sub-Microsecond Latency: Scale-Up Networking via Tomahawk 5 Ultra
To optimize inter-GPU scale-up communications within tightly coupled computing domains , UfiSpace introduces the S9322-64EO data center switch. Powered by Broadcom's Tomahawk 5 Ultra silicon, the 2RU platform provides 51.2 Tbps aggregate switching bandwidth across 64x 800G OSFP ports. 

Engineered specifically as an open Scale-Up Ethernet architecture for inter-rack expansion across ultra-large GPU clusters, the S9322-64EO addresses the performance bottlenecks of small-packet collective communications (such as All-Reduce and All-to-All). By deploying Scale-Up Ethernet (SUE) with AFH Gen 2 header compression, compressing conventional protocol overhead from approximately 46 bytes down to 10 bytes, and delivering deterministic latency below 250 nanoseconds, the platform dramatically maximizes payload efficiency. This provides cloud operators with a scalable, open Ethernet solution for multi-node GPU scale-up fabrics across racks.

"Scaling multi-trillion parameter AI clusters requires open, high-density infrastructure that delivers greater architectural flexibility while addressing severe thermal and latency constraints," said Vincent Ho, CEO of UfiSpace. "By integrating Broadcom's Tomahawk 6 and Tomahawk 5 Ultra silicon into our disaggregated 1.6T switching platforms, UfiSpace empowers cloud providers with field-proven, energy-efficient architectures engineered to maximize GPU utilization and optimize TCO across both Scale-Out and Scale-Up AI fabrics."

Deep-Buffered AI Scale-Across DCI (Qumran4D / Qumran3D)
As modern AI architectures shift toward Geo-Distributed Training and multi-region RAG/Inference clusters, cross-datacenter links are vulnerable to massive traffic bursts during multi-terabyte LLM Checkpoint synchronization and real-time dataset ingestion. To resolve these WAN constraints, UfiSpace highlights its open router family powered by Broadcom Qumran architecture:

• S9630-32HO (51.2T Open Router): Driven by Broadcom Qumran4D with an embedded 32GB HBM Deep Buffer, line-rate MACsec/IPsec encryption, and OpenZR+ coherent optics, this 3RU router delivers zero-packet-loss DCI routing across distances over 400 km without requiring external optical transponders . 
•  S9620-32E (25.6T Metro & Edge Router): Powered by Broadcom Qumran3D in a 2RU chassis, this router provides 32x 800G QSFP-DD ports backed by a 32GB Deep Buffer. It provides operators with a cost-effective, deep-buffered edge solution for regional DCI and metro core upgrades.

To demonstrate real-world commercial readiness, UfiSpace's Qumran4D-powered S9630-32HO open router platform will be showcased at Broadcom's booth during OCP APAC 2026.

About UfiSpace
UfiSpace is a technology leader in disaggregated open networking solutions, empowering telecom operators, cloud service providers, and hyperscale data centers to modernize their network infrastructure. Combining telecom-grade engineering with deep industry expertise, UfiSpace delivers high-performance, energy-efficient switch and router platforms designed to meet the demands of AI scale-out/scale-up fabrics and multi-terabit networks. As an active contributor to the Open Compute Project (OCP), UfiSpace collaborates with a global partner ecosystem to advance open standards, optimize Total Cost of Ownership (TCO), and accelerate the worldwide adoption of disaggregated architectures. Learn more at our website.
Follow us on our social media channels: LinkedIn, Facebook, and YouTube

MEDIA CONTACT
Name: Betty Wu     
E-mail: betty.wu@ufispace.com