ZKKR Unveils AS6208-A8V4P AI Training & Inference Server

Global AI infrastructure investment is scaling at an unprecedented pace. According to IDC, the worldwide server market reached $122.6 billion in vendor revenue in Q1 2026 — a 30.4% year-over-year increase, with GPU-accelerated servers generating $68.9 billion, representing 56.2% of total market revenue. TrendForce projects global AI server shipments to grow over 28% year-over-year in 2026, with high-end AI training servers accounting for approximately 55% of total shipments. Meanwhile, AI inference is rapidly gaining momentum: the top five North American CSPs are expected to see total AI inference compute capacity grow by approximately 122% in 2026 — more than double the growth rate of training capacity. The training-inference integrated architecture is emerging as the dominant data center paradigm.

At the GPU level, the NVIDIA H200 — built on the Hopper architecture — is the first GPU to offer 141GB of HBM3e memory at 4.8 TB/s, delivering nearly double the capacity of the H100 with 1.4× more memory bandwidth. Its FP16 compute reaches 990 TFLOPS, approximately twice the performance of the H100. To meet the stringent demands that high-performance GPUs like the H200 place on server systems — in cooling, power delivery, and interconnect bandwidth — ZKKR introduces the AS6208-A8V4P, a 6U 2-socket 8-GPU AI training and inference server. From architecture design to hardware layout, every aspect is optimized around passive-cooled professional GPUs such as the NVIDIA H200 and PRO6000, delivering a robust compute foundation for billion-parameter model training and millisecond-latency inference.

CPU-GPU Heterogeneous Synergy for Maximum Compute Power. 

The AS6208-A8V4P is powered by dual AMD EPYC™ 9004/9005 series processors, featuring advanced process technology and multi-core architecture with support for up to 500W TDP, delivering exceptional general-purpose compute capability for complex task scheduling and data preprocessing. Paired with 8 high-performance 600W GPU cards, the system achieves precise thermal control for passive-cooled professional GPUs including the NVIDIA H200 and PRO6000. For AI inference, it delivers millisecond-latency intelligent decision-making at scale; for large model training, it accelerates parameter iteration of complex Transformer-based models, dramatically reducing training cycles.

PCIe 5.0 High-Speed Interconnect Eliminates Bandwidth Bottlenecks. 

With PCIe 5.0 technology delivering 32Gbps transfer rates — double that of the previous generation — the AS6208-A8V4P establishes an ultra-high-speed data channel between CPU and GPU, ensuring rapid, stable data flow for large-scale AI training and compute workloads while eliminating performance bottlenecks caused by insufficient bandwidth.

Enterprise-Grade Reliability for 24/7 Uninterrupted Operation. 

The system supports N+N redundant power supplies with dual-grid connectivity, ensuring uninterrupted operation even during sudden power failures. Intelligent power capping dynamically maintains the server within optimal power ranges, while N+1 redundant fans ensure continuous operation even in high-temperature environments. The onboard AST2600 BMC management chip supports IPMI2.0, KVM Over IP, SOL, and SNMP for remote management, enabling data center operators to reduce costs and improve operational efficiency.

Compute is the engine; storage is the foundation. Guided by its "All-in-AI" strategy and backed by integrated R&D, design, manufacturing, and service capabilities, ZKKR continues to build a full-scenario compute and storage server portfolio spanning training, inference, and storage. For more information about ZKKR solutions, please contact our sales representatives. ZKKR looks forward to partnering with you to shape the future of the data era.