The Cisco UCS C885A M8 Rack Server is a dense-GPU server designed to deliver massive, scalable accelerated compute capabilities to address the most demanding AI workloads, including deep learning/Large Language Model (LLM) training, model fine-tuning, large model inferencing, and Retrieval-Augmented Generation (RAG).

To deliver massive accelerated compute performance in a single server, the server offers a choice of eight GPUs of the following types:

• NVIDIA® H100 SXM or NVIDIA® H200 Server PCI Express Module (SXM) GPUs. SXM is a socket-based GPU interconnect method used by NVIDIA GPUs.

• AMD MI300X OCP Accelerator Model (OAM) GPUs. OAM is an Open Compute GPU interconnect standard that avoids GPU vendor lock in.

For north-south traffic, the server supports up to two NVIDIA BlueField-3 B3220 DPUs or up to two NVIDIA ConnectX-7 NICs.

For east-west traffic, the server supports up to eight NVIDIA ConnectX-7s or BlueField-3 B3140H SuperNICs for GPUs enabling AI model training across a cluster of dense-GPU servers.

The server is offered in fixed configurations with resources optimized for intensive AI and HPC workloads.

Server Front and Rear Views

The front of the server contains fans, storage modules, system LEDs, and USB port access covered by a bezel.

Slot Numbering, Server Front

Slot numbers are silk screened onto the front chassis. You can use the silkscreening on the chassis to locate appropriate slots.

Server Rear View

The rear of the server contains compute and networking, as well as power supplies and management ports.

Slot Numbering, Server Rear View

Slot numbers are silk screened onto the rear of the chassis. You can use the silkscreening on the chassis to locate appropriate slots.

Server Configurations

Cisco UCS C885A M8 Rack Server is shipped with a few components pre-configured/auto included, but for a few components, you will need to make a selection.

Each UCSC-885A-M8-H11 server supports the following:

Each UCSC-885A-M8-H12 server supports the following:

Each UCSC-885A-M8-H13 server supports the following:

Each UCSC-885A-M8-H20 server supports the following:

Each UCSC-885A-M8-H21 server supports the following:

Each UCSC-885A-M8-H22 server supports the following:

Each UCSC-885A-M8-H23 server supports the following:

Each UCSC-885A-M8-H25 server supports the following:

Each UCSC-885A-M8-H26 server supports the following:

Each UCSC-885A-M8-H27 server supports the following:

Each UCSC-885A-M8-H28 server supports the following:

Each UCSC-885A-M8-H29 server supports the following:

Each UCSC-885A-M8-M3X0 server supports the following:

Each UCSC-885A-M8-M3X1 server supports the following:

Each UCSC-885A-M8-M3X2 server supports the following:

Each UCSC-885A-M8-M3X3 server supports the following:

Each UCSC-885A-M8-M3X4 server supports the following:

Each UCSC-885A-M8-M3X5 server supports the following:

Each UCSC-885A-M8-M3X6 server supports the following:

CPU Tray

The CPU Tray contains the compute and network functions of the server.

DC-SCM Module

The DC-SCM module is supported on the CPU Tray. The module provides an interface to the server, and enables enforcing a common, robust security module across the servers and its connected peers that support the security model. By default, the DC-SCM module is the primary management interface for the server, but the OCP NIC can also be used for the management interface if desired.

Cisco M.2 NVMe BOOT RAID Controller

The server supports one internal boot-optimized RAID card, the Cisco M.2 NVMe BOOT RAID Controller (UCS-M2-NVRAID).This card is a standard half-height, half-length PCIe x8 NVMe RAID adapter card, which features two 960 GB NVMe M.2 SSDs with RAID 0/1 functionality.

• The card features host interface support of PCIe Gen4 x8, and for each M.2, up to PCIe Gen4 x4

• Each M.2 NVMe SSD can be either 110 mm and 80mm in length.

• Each M.2 NVMe SSD has 960 GB of storage capacity.

• Each SSD is uniquely numbered:

  • Drive 251 is the SSD at the "top" of the module. This drive is located closest to the right-angle bend in the module faceplate.

  • Drive 252 is the SSD at the "bottom" of the module. This drive is located closest to the golden fingers connectors that insert into the PCIe socket.

The card is a boot-optimized RAID card, so it controls striping and mirroring (RAID 0/1) of data for server boot up, OS loading, and logging. Because the card is boot-optimized, it does not control user data I/O, such as reads and writes to the server’s SAS and NVMe storage drives.

The card installs into slot 15 as a hardware option but is not a field-replaceable unit (FRU). If you want the Cisco M.2 NVMe BOOT RAID Controller installed in your server, it will be pre-installed in a new CPU Tray, which is a FRU. Contact Cisco to order a new CPU Tray and specify all modules and 12V PSUs installed on the CPU Tray so that you can receive an exact replacement.

Fan and Storage Locations

The server contains 12 fans for proper cooling as well as three groups of hot-swappable NVMe SSDs for storage.

Front Panel LED Indicators

Rear Panel LED Indicators

Note	This LED is hardware and software controlled. See POST Error Codes for a detailed description of seven-segment LED Codes.

Server Components

The server has the following main components. Some sub components are contained within specific trays in the server. For example, the CPUs, heatsinks, and DIMMs are contained on the CPU tray. If these components need to be replaced, you will order a replacement CPU tray, and replace the entire tray.

Mainboard Components

The server mainboard components are documented here for reference. These components are not serviceable. If service is required for any of these components, you will order a new CPU tray.

Note	Despite the CPU label that shows CPU replacement, individual CPUs, heatsinks, and so on, are not field replaceable.

DC-SCM Board Components

The DC-SCM board has three available variants as shown in the following table.

The DC-SCM PCB components are documented here for reference. These components are not serviceable. If service is required for any of these components, you will order a new DC-SCM tray.

Power Distribution Board Components

The power distribution board receives input facility power from the server's PSUs and conditions it for use by the different subsystems in the server.

The power distribution board components are documented here for reference. These components are not serviceable.

Mid-Plane Board Components (Front)

The mid-plane board connects all relevant subsystems in the server, such as networking, compute, storage, and so on.

The mid-plane board's front components are documented here for reference. This board and its components are not serviceable.

Mid-Plane Board Components (Back)

The mid-plane board's rear components are documented here for reference. This board and its components are not serviceable.

HIB Board Components

The host interface board (HIB) and its components are documented here for reference. This board and its components are not serviceable.

FIO Board Components

The FIO board and its components are documented here for reference. This board and its components are not serviceable.

CEM 8X BB Board Components

The CEM 8x BB board and its components are documented here for reference. This board and its components are not serviceable.

Left CEM Riser Board Components

The CEM riser comes in both left and right versions based on where they mount into the chassis. Each CEM riser is different from the other. The left CEM riser is a three-slot metal cage that contains the left CEM board.

The left CEM riser, the left CEM board, and its components are documented here for reference. The riser, board, and its components are not serviceable.

Right CEM Riser Board Components

The CEM riser comes in both left and right versions based on where they mount into the chassis. Each CEM riser is different from the other. The right CEM riser is a two-slot metal cage that contains the right CEM board.

The right CEM riser, the right CEM board, and its components are documented here for reference. The riser, board, and its components are not serviceable.

Fan Extension Board Components

The fan extension board contains the fan interposer board, which connects to the server midplane. The fan extension board and its components are documented here for reference. This board and its components are not serviceable.

Fan Interposer Board Components

The fan interposer card provides connection between the server midplane and a fan to provide power and signaling between the server and the individual fans.

The fan interposer board and its components are documented here for reference. This board and its components are not serviceable.

SAS IO Board Components

The SAS I/O board and its components are documented here for reference. This board and its components are not serviceable.

Dual M.2 Board Components

The dual M.2 board contains two NVMe SSDs that are only for server boot up. The Dual M.2 board and its components are documented here for reference. This board and its components are not serviceable.

Fan and Storage Locations

The server has a fan module consisting of twelve individual fans that provide front-to-back airflow for optimal cooling.

Below the fan module, the server has a storage subsystem that supports up to 16 hot swappable NVMe SSDs and eight host swappable SAS SSDs for customer data.

DIMM Slot Locations

The server contains two CPUs (0 and 1) each with two banks of six DIMMs channels. CPU 0 has DIMM channels A through F and G though L for a total of 12 DIMMs.

CPU 1 has DIMM channels A through F and G through L for a total of 12 DIMMs.

The individual DIMMs are not serviceable.

DIMM slots are labeled sequentially (A through L) for each CPU. Use the following recommendations for memory channels per configuration.

Memory Module Population Rules

All DIMMs must be DDR5 DIMMs. 16GB, 24GB, and 32GB density DRAMs are supported. The following DIMM types and bit widths are supported:

• DIMM capacities from 16GB to 1TB depending on type and configuration

• RDIMM: 1Rx8, 2Rx8, 1Rx4, and 2Rx4

• 3DS RDIMM: 2Rx4 2H and 2Rx4 4H o 2Rx4 8H, which support capacities of 512GB/768GB/1024GB for 16Gb/24Gb/32Gb die densities, respectively.

• 80b x4, 80b x8, and 72b x4 ECC support

• 80b x4 AMDC, 80b x8 and 72b x4 optimized Bounded Fault ECC DRAM support

The following general population restrictions apply:

• Do not mix DIMM module types within a memory channel. All must be RDIMM or RDIMM 3DS module types with the same ECC configuration.

• Do not mix x4 and x8 DIMMs within a memory channel.

• Do not mix 3DS and non-3DS memory modules in a 2DPC system.

General population recommendations:

• Mixing DIMMs of different ranks (i.e., 1R, 2R, 3DS) across memory controller channels will change or disable the chip select interleaving feature and may have performance impact. Mixing DIMMs of different ranks within a memory controller channel is not supported.

• Mixing DIMMs from different manufacturers is supported across memory controller channels, but is not supported within a memory controller channel.

• Mixing DIMMs of different density (i.e., 16Gb, 24Gb, and 32Gb DRAMs) within a memory controller channel is not supported. Mixing DIMM density across memory controller channels is supported, though a performance impact may be noted. Platform developers are expected to validate configurations that mix DIMM density across memory controller channels.

Redundant PSU Locations

The server has PSUs rated at 12V and 54V. Different PSU types support different power redundancy. Full grid redundancy is not supported.

Support for PSU Redundancy and Continuous Operation

The system includes eight power supply units (PSU) configured for 12V CRPS 1+1 and 54V PSU 4+2 redundancy. Refer to the following additional considerations:

• For 12V CRPS 1+1

  • If one PSU fails, troubleshoot the cause and replace the failing PSU immediately. If one PSU fails, the server continues to operate at full performance, and the PSU statues/health LED glows solid amber.

• For 54V AC PSU 4+2

  • If only two 54V PSUs are installed, the system can power on and operate without the GPU Tray installed.

    Note	This configuration is not a standard runtime config. It is used for system debug purposes only.

  • If only three 54V AC PSUs are installed, the server operates, but with performance reduced performance (power capping every GPU to 60%).

  • If four or more 54V AC PSUs are installed - the servver operates at full performance.