Specifications
Form factor
4U rackmount, dual-processor
CPU
Up to 2x AMD EPYC 9004 series, up to 128 cores each, up to 400 W TDP
GPU modules
8-GPU module, 4-slot direct-connect module, 8-slot direct-connect module
Max double-width GPUs
10 (8-GPU module and 8-slot direct connect); 6 (4-slot direct connect)
GPU topologies
Five selectable topologies with one-click configuration switching
CPU to GPU ratios
1:4 and 1:8
Memory slots
Up to 24x DDR5 up to 5600 MT/s, RDIMM and LRDIMM
Memory capacity
Up to 6 TB across 12 memory channels
CXL support
CXL 1.1
Drive bays
Up to 24x 2.5/3.5-inch SAS/SATA/NVMe plus 2 onboard M.2
NVMe support
Up to 8 NVMe drives per chassis
Storage design
Up to 400 TB per system
PCIe — 8-GPU module
8x double-width PCIe 5.0 x16 plus 4x single-width PCIe 5.0 x16
PCIe — 8-slot direct connect
10x double-width PCIe 5.0 x16
Dedicated slots
1x OCP 3.0 network adapter slot, 1x storage controller slot
Networking
1/10/25/50/100/200/400GE adapters
RAID
RAID 0/1/10/5/6/50/60 with Tri-Mode support
Fans
8x 8056 fans, N+1 redundancy, hot swap
Power supplies
4x 2000 W / 2400 W / 2700 W Platinum or Titanium, N+N redundancy
Management
HDM agentless management, H3C UniSystem, LCD touch module, U-Center
Security
TPM/TCM modules, chassis intrusion detection, dedicated security processors
Dimensions
174.8 x 447 x 828 mm (4U, without security bezel)
Operating temperature
5 C to 35 C, configuration dependent
Overview
The H3C UniServer R5350 G6 is a 4U dual-socket AI server built on two AMD EPYC 9004 series processors and a choice of three accelerator modules: an 8-GPU module, a 4-slot direct-connect module, or an 8-slot direct-connect module. Depending on the module selected, the chassis holds up to ten double-width GPUs with one-click switching between five predefined topologies.
Because the same platform covers 1:4 and 1:8 CPU-to-GPU ratios, it can be re-provisioned between inference-heavy and training-heavy fleets without a chassis change. 24 DDR5 slots across 12 channels reach 6 TB with CXL 1.1 support, and the storage side is integrated rather than bolted on: up to 24 front bays, eight of them NVMe, plus two onboard M.2 modules and a dedicated storage controller slot, for a 400 TB per-system design.
Expansion runs to twelve PCIe 5.0 slots in the 8-GPU configuration with a dedicated OCP 3.0 network adapter slot and support for 400GE adapters, so the server can host a full AI fabric as well as compute. H3C manages the platform through HDM agentless management, H3C UniSystem and the U-Center data centre platform, with TPM/TCM security modules and chassis intrusion detection.
Key Benefits
One chassis, two fleet profiles — five GPU topologies and 1:4 or 1:8 accelerator ratios support both training and inference from the same SKU. Compute-storage integration — up to 10 double-width GPUs alongside a 400 TB storage design simplifies large-dataset AI deployments. Current-generation interconnect — PCIe 5.0, CXL 1.1 and 400GE networking keep the platform aligned with new fabric and memory tiers. Serviceability — N+1 hot-swap fans and N+N power supplies across 2000 W to 2700 W Platinum or Titanium units.
Applications
Deep learning model training and inference, HPC and data analytics clusters, H3C/HGX-based AI pods, memory-heavy large-model serving that benefits from CXL and 6 TB of DDR5, and enterprise or carrier AI infrastructure requiring 400GE east-west networking.
Request a Quote — H3C UNISERVER R5350 G6 — 10-GPU DUAL EPYC 9004 AI SERVER
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →