INNOSILICON Unveils Full Suite of High-Speed UALink IP, Powering Next-Gen Token Factories with Enhanced Productivity

June 04, 2026  -- INNOSILICON Technology today officially announced the launch of its comprehensive high-speed UALink (Ultra Accelerator Link) solution portfolio, featuring industry-leading single-lane 112G/224G high-performance PHY technology. Focused on next-generation AI Scale-Up core technologies for hyperscale compute supernode expansion, the solution delivers a major breakthrough in AI infrastructure by eliminating interconnect barriers between AI systems, deeply integrating the core capabilities of Token factories, and significantly increasing Token production efficiency.

Technology First: Building a Solid Foundation for Token Factory Capacity

As large language model parameters grow larger and AI architectures evolve toward reinforcement learning, reasoning, AI agents, and Prefill-Decode (PD) disaggregation, the performance ceiling of individual chips has become increasingly apparent. The industry is rapidly shifting toward high-speed interconnect technologies to link multiple compute devices and enable scalable compute aggregation. Meanwhile, mainstream algorithms including Tensor Parallelism (TP) and Expert Parallelism (EP) place stringent demands on inter-chip bandwidth and latency.

Designed for AI Scale-Up and thousand-GPU AI clusters, UALink is an open high-speed interconnect protocol jointly developed by over 70 global technology companies including AMD, Google, Microsoft, and Alibaba, with INNOSILICON among them. Unlike conventional Ethernet architectures, UALink eliminates the overhead of redundant TCP/IP networking stacks and adopts a streamlined four-layer hardware protocol stack with native memory read/write semantics, enabling accelerators to directly access remote GPU memory across chips without memory copying.

With a fixed FLIT frame structure and a hardware-level retransmission mechanism, UALink delivers ultra-low latency, high determinism, and high bandwidth utilization, achieving over 93% efficiency for small packet transmission. By reusing existing Ethernet PHYs, cables, and optical modules, UALink strikes an optimal balance between an open ecosystem and cost efficiency. Unlike InfiniBand, which is closed and costly, and unlike traditional Ethernet, which suffers from high latency, excessive overhead, and significant jitter, UALink offers a superior alternative. It stands as one of the leading interconnect solutions for Token factories and large-scale model training clusters.

The newly released UALink IP suite from INNOSILICON Technology has successfully passed customer system validation and supports all types of XPU chips. Widely deployed in AI super nodes, the suite is compatible with 800G/1.6T optical modules as well as INNOSILICON and CPO optoelectronic convergence architectures. Adhering to a differentiated R&D strategy, INNOSILICON Technology has embedded 224G core technologies into its mature 112G products currently in mass production sampling. The solution supports channels with up to 40dB loss, achieves a raw bit error rate (BER) of 2e-10 in typical optical module application scenarios, and maintains stable performance across full PVT (process, voltage, temperature) variations. Both transmitter and receiver have fully passed protocol compliance testing. Equipped with advanced error correction and low-power design, the IP supports seamless product upgrades and reduces performance loss caused by transmission latency.

Explore UALink IP

INNOSILICON Technology's 224G UALink has successfully completed tape-out, with all simulation and verification metrics meeting design targets. The company plans to initiate global licensing within the year. This technology will further enhance optoelectronic interconnection capabilities, establish end-to-end high-speed channels for the entire Token production and transmission flow, and address the industry-wide challenges of fragmented computing resources and constrained data throughput. Leveraging the open ecosystem advantages of UALink—which ensures compatibility with standard Ethernet hardware—the solution significantly reduces deployment costs compared to closed, proprietary protocols and enables seamless integration of computing devices from multiple vendors.

Lead the world. Redefine next-gen AI infrastructure

Leveraging nearly twenty years of technical excellence, Innosilicon has developed an AI "three-piece suite" of storage, compute, and interconnect. The portfolio features high-end memory interfaces ensuring high-speed data access, 112G/224G UALink eliminating transmission bottlenecks, and the self-developed Fantasy series GPUs delivering robust computing power.

The synergistic effect of these core technologies fully empowers AI computing clusters and the Token factory ecosystem, driving deep adoption of compute-storage integration and optoelectronic convergence, breaking down device interconnect barriers, and enabling seamless, end-to-end connectivity across Token storage, production, and transmission. In addition, UALink features a built-in hardware encryption module supporting confidential computing, meeting the security requirements for sensitive data training in finance, government, and other sectors.

With proven mass production and delivery capabilities, Innosilicon has served over 300 global enterprises, with cumulative shipments of self-developed IP chips exceeding 10 billion units, establishing itself as a core partner and supplier to global Token factories. Going forward, Innosilicon will continue to deepen its focus on high-speed interconnect, further optimize UALink technology, enable large-scale upgrades of AI computing clusters, and help global clients seize the strategic high ground in AI infrastructure.