GPU Density and Interconnection 8 x NVIDIA HGX H100/H200 700W SXM5 (or optional 8 x A100 500W SXM4), fully interconnected through NVLink, GPU-GPU bandwidth up to 900 GB/s, 1.5 times higher than the previous generation Support NVIDIA MIG multi instance GPU, which can virtually create 7 instances in one card, achieving multi tenant isolation CPU and Memory 2 x 4th/5th generation Intel Xeon Scalable (such as Platinum 8480+), with a maximum of 56 cores per chip, TDP 350W 32 × DDR5-4800 DIMM, maximum 4TB, bandwidth increased by 50% compared to DDR4, meeting the caching requirements of large model parameters I/O and Storage Front window: 8 × 2.5 "NVMe/SAS/SATA+16 × EDSFF E3. S Gen5 NVMe (optional 122.88TB full flash), supports NVIDIA GPUDirect Storage, DMA between GPU and disk bypasses CPU to reduce latency Motherboard: 10 x PCIe 5.0 slots (x16/x8), can accommodate 400GbE, InfiniBand NDR, DPU, etc; Dual OCP 3.0 network card slots, flexible replacement of 25/40/100/200 GbE Heat dissipation and power supply 6U cellular duct+6 sets of hot swappable redundant fans, supporting long-term operation in the 35-40 ° C computer room 4 × 2800W 80Plus titanium hot swappable power supply (2+2 redundancy), supporting AC 200-240V or HVDC, with a full load power consumption of approximately 10kW Management and Security iDRAC9 Datacenter， Provide RESTful API, telemetry CloudIQ AIOps； Support zero trust architecture: silicon trust root, signed firmware TPM 2.0、 multi-factor authentication typical scenario Training/fine-tuning of billions of parameter large models, generation of AIGC content, development of deep learning, HPC computational fluid dynamics, digital twins, metaverse rendering
Indiquez la référence, la quantité et la date souhaitée. Nous confirmons stock ou délai, proposons une alternative si besoin et émettons une facture commerciale avec code SH.
Envoyez votre nomenclature ; réponse sous un jour ouvré.
Nous contacter
