Specifications
Form Factor
8RU rack server, fixed configurations
GPU Count
8x AMD Instinct MI350X OAM (UBB 8-GPU baseboard)
GPU Interconnect
8-way OAM module with AMD Infinity Fabric links
Processors
2x 5th Gen AMD EPYC 9575F — 64 cores, 3.3 GHz base / 5.0 GHz boost
Memory
24x DDR5 up to 6,400 MT/s RDIMM — 96 GB or 128 GB modules, up to 3 TB
East-West Networking
8x PCIe Gen5 x16 HHHL — NVIDIA ConnectX-7 1x400G or AMD Pensando Pollara 400 1x400G
North-South Networking
5x PCIe Gen5 x16 FHHL — ConnectX-7 2x200G, 1x400G or 4x25G
DPU / SuperNIC
NVIDIA BlueField-3 DPU or AMD Pensando Pollara NIC per GPU for data acceleration
OCP Slot
1x OCP 3.0 PCIe Gen5 x8 — Intel X710-T2L 2x10G RJ45
Boot Storage
Up to 2x 960 GB M.2 NVMe SSD
Internal Storage
Up to 2x 2.5 inch U.2 NVMe SSD for data caching
CPU Power
2x 2,700 W 80 PLUS 12 V CRPS, N+1 redundant hot-swap
GPU Power
6x 3,000 W 80 PLUS 54 V MCRPS, N+2 redundant hot-swappable
Cooling
Air-cooled with high-static-pressure fan wall and GPU/CPU sled isolation
Management
Cisco Intersight with Device Connector inventory plus local CIMC / Redfish
Ordering Part Number
UCSC-885A-M8-M352 (8x MI350X, 24x DDR5, 8x 400G East-West)
Alternative GPU Build
Also offered as 8x NVIDIA HGX H100 or H200 Tensor Core GPUs per configuration
Target Workloads
LLM training and fine-tuning, large-model inference, RAG pipelines, HPC
Overview
The Cisco UCS C885A M8 is a density-optimised 8RU GPU server built around the NVIDIA HGX platform architecture and populated with eight AMD Instinct MI350X OAM accelerators. Each GPU is paired with an NVIDIA ConnectX-7 400G NIC or an AMD Pensando Pollara 400 NIC, so east-west fabric bandwidth scales with compute rather than becoming the bottleneck when training runs are spread across a cluster of nodes.
Compute is driven by two 5th Gen AMD EPYC 9575F processors running 64 cores each at up to 5.0 GHz boost, with 24 DDR5-6400 RDIMM slots delivering up to 3 TB of system memory. Local storage is deliberately small and fast — M.2 NVMe boot plus two U.2 NVMe caching drives — because the platform is designed to stream training data from a fabric-attached or disaggregated storage tier.
Power delivery is split by rail the way current HGX-class designs require: 12 V CRPS supplies for the CPU complex and 54 V MCRPS supplies for the GPU baseboard, both hot-swappable and redundant. Cisco Intersight provides inventory and lifecycle management, with local CIMC/Redfish available for out-of-band operations during rollout.
Key Benefits
8-way MI350X density in 8RU: eight OAM accelerators on a single UBB baseboard deliver rack-scale training throughput without a 10U footprint. Fabric-matched east-west I/O: eight 400G NICs (ConnectX-7 or Pensando Pollara) keep GPU-to-GPU cluster traffic balanced. Memory headroom: up to 3 TB DDR5-6400 across 24 RDIMMs sustains large optimiser states and long-context KV cache. Rail-split redundant power: independent 12 V and 54 V supplies with N+1 / N+2 redundancy protect AI training jobs from single-PSU failure. Fixed configurations: factory-built SKUs remove configuration drift and shorten quote-to-deploy time for volume AI cluster orders.
Applications
Designed for AI research labs, neoclouds and enterprise AI factories running large language model pretraining and fine-tuning, retrieval-augmented generation at scale, recommendation and ranking model training, molecular and scientific HPC simulation, and multi-node GPU cluster deployments where 400G east-west networking is mandatory.
Request a Quote — CISCO UCS C885A M8 — 8RU 8-GPU AMD INSTINCT MI350X 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 →