토탈 IT 솔루션 리더 슈퍼마이크로(Supermicro)가 4세대 AMD CDNA 아키텍처 기반 ‘AMD 인스팅트 MI350’ 시리즈 GPU를 탑재한 H14세대 AI 솔루션을 공개하며, 대규모 AI 학습과 추론 워크로드에 최적화된 성능과 유연성을 제공한다.
Supports both water cooling and air cooling, maximizes performance and efficiency with 2.3TB HBM3e memory per server

Supermicro, a total IT solutions leader, is introducing new AI solutions featuring the AMD Instinct MI350, delivering optimized performance and flexibility for large-scale AI training and inference workloads.
On the 19th, Supermicro unveiled its H14 generation AI solution equipped with the 'AMD Instinct MI350' series GPU based on the 4th generation AMD CDNA architecture.
This new product supports both liquid-cooled (DLC) and air-cooled options and provides performance and flexibility optimized for large-scale AI training and inference workloads with up to 2.3TB of HBM3e large memory per server.
The new H14 GPU-optimized solution combines dual AMD EPYC 9005 series CPUs and AMD Instinct MI350 GPUs.
The 288GB HBM3e provided per GPU is 1.5 times larger than the previous generation, the memory bandwidth is 8TB/s, and the petaflops performance based on FP16 and FP8 is 1.8 times larger than the previous generation.
Based on this, AI models as well as complex scientific simulations can be processed quickly.
The water-cooled 4U system is a DL system developed by Supermicro.By applying the C (Direct Liquid Cooling) architecture, it efficiently controls heat in high-density environments and reduces power consumption by up to 40%.
An air-cooled 8U system is also available, allowing flexible rack scale construction to suit your data center environment and budget.
Both options support the industry-standard open multi-architecture OCP Accelerator Module (OAM) to ensure compatibility.
“Supermicro’s Data Center Building Block Solutions (DCBBS) are a proven platform for rapidly delivering end-to-end AI and HPC systems,” said Charles Liang, president and CEO of Supermicro. “With the addition of the AMD Instinct MI350 series, our H14 generation solutions enable customers to efficiently and economically design and build their next-generation data centers.”
“The Instinct MI350 series delivers up to 40 percent greater token throughput per dollar while maintaining an industry-standard form factor, and when combined with Supermicro’s integrated air- and water-cooled platform, provides flexibility and energy efficiency,” said Lisa Su, AMD president and CEO.
Paul Schell, industry analyst at ABI Research, said, “Supermicro’s introduction of an expanded memory solution that supports both air and liquid cooling reflects the growing demand for high-performance, low-power infrastructure that spans both AI training and inference. It will enable cloud service providers (CSPs) and enterprises to reliably operate large-scale AI workloads.”