- Server
- Components
- Carbon Fibre
- Analog to digital conversion chip
- power management IC
- storage
- H3C router
- H3C server
- Inspur server
- hard disk
- HART Loop Converter Analog Input
- reader
- Ethernet module
- EDGE Guardmaster
- MAT Guardmaster
- MatGuard Mat Manager
- relay
- power control
- switch
- servo drive
- I/O module
- Bridge connector (jumper plug)
- Human Machine Interface (HMI)
- power monitor
- 1420 power monitor
- 1411 current transformer
- socket
- system
- electric relay
- frequency converter
- Install the GND connector on the board
- transceiver
- Module Option Card
- module
0102030405
NF5468A5
product detail
heterogeneous computing power
Up to 8 x dual width PCIe Gen4 acceleration cards in 4U: verified support for NVIDIA A100/H100, AMD Instinct, Intel Habana, Cambrian MLU, Suiyuan, Inspur self-developed M10A VPU, etc
2 x AMD EPYC 7003 (Milan/Milan-X), up to 64 cores/280 W per chip, 128 cores/256 threads, 1536 MB L3 Cache; 32 x DDR4 DIMM, maximum 8 TB, memory bandwidth 750 GB/s
CPU-GPU direct connection topology (without PCIe Switch), latency reduced by 200-300 ns, single CPU ? GPU bandwidth of 128 GB/s, 4 times higher than traditional solutions
Storage and IO
Front window 12 × 3.5 ″/2.5 ″ hot swappable (SAS/SATA/NVMe hybrid)+2 × M.2 NVMe mirror boot
Rear window 2 x 2.5 "+1 x OCP 3.0 network card (hot swappable, 1 min replacement)+3 x single width PCIe x16 (25/100/200 GbE or RAID card)
Optional 8/16 port 12 Gb/s RAID card, supporting RAID 0/1/5/6/10/50/60
Power supply and heat dissipation
4 x 2200 W 80 PLUS Platinum (1600-3000 W range), supporting 3+1 or 2+2 redundancy
Mid mounted "fan wall" (6 sets of sub fan modules)+partitioned air guide cover, forming CPU/GPU independent air ducts; Can stably load 2 × 280 W CPUs+8 × 300 W GPUs, with a temperature of approximately 60 ℃ when the GPU continues to operate at 330-340 W
Performance testing (official prototype: 2 x EPYC 7543 32C+512 GB DDR4+8 x A100 40 GB)
HPL CPU Floating Point: 4.2 TFLOPS (Double Precision)
Memory bandwidth: 373 GB/s (efficiency 91%)
ResNet50 training: 21 486 images/s (8 × A100)
ResNet50 inference: 13% increase compared to T4 nominal value
Video transcoding: M10A VPU single card ETHASH 108 MH/s, RTX3090 350 W continuous 60 ℃
typical scenario
Billion parameter large model training/fine-tuning, CV/NLP/speech recognition, video encoding and decoding, blockchain HPC、 Cloud gaming, metaverse rendering