By AMD EPYC™ 9005 CPUs, AMD Instinct™ MI350 Series GPUs, and AMD Pensando™ Pollara 400 AI NIC
Building enterprise-grade AI requires deep compatibility between your compute, acceleration, and networking layers. By standardizing these core elements on an integrated AMD ecosystem, ASRock Rack systems remove common data bottlenecks. This unified approach simplifies cluster management while ensuring your infrastructure delivers the consistent throughput required for complex, real-world AI training.
Powered by the advanced 4th Gen AMD CDNA™ architecture, AMD Instinct™ GPUs are engineered to accelerate generative AI and large-scale LLM training. Equipped with up to 8TB/s of bandwidth and expanded datatype support (including MXFP4 and MXFP6) to maximize throughput, these accelerators are supported by the fully open-source AMD ROCm™ software stack to power your most demanding inference, training, and HPC workloads.
Leveraging the 4th Gen AMD CDNA™ architecture, the AMD Instinct™ MI355X GPU maximizes AI and HPC performance for high-density environments, featuring 288GB HBM3E, and 8TB/s bandwidth.
Designed for seamless integration into existing infrastructure, the AMD Instinct™ MI350P PCIe® card scales generative and agentic AI workloads, delivering high performance, cost efficiency, and an open software stack.
Based on the “Zen 5” architecture and offering up to 192 cores, the AMD EPYC ™ 9005 Series Processors handle heavy dataset preprocessing and AI orchestration tasks with ease. Their high-bandwidth data channels keep your GPU accelerators fed with information, ensuring the entire system operates at its maximum potential.
Networking often becomes the hidden bottleneck in AI clusters. As the first UEC-ready AI NIC, the Pollara 400 delivers 400Gbps throughput and uses intelligent packet spraying and programmable RDMA to prevent congestion. It acts as a direct data pipeline, keeping your GPUs supplied with the information they need to avoid idle cycles.
