Supermicro GPU SuperServer 420GP-TNAR+

Description

Along with 6x hybrid drive bays in front, there are 4x more drive bays in back dedicated for NVMe storage devices. The Supermicro GPU SuperServer 420GP-TNAR+ is designed for applications in high-performance computing (HPC), AI and Deep Learning, to name a few. It is specifically designed to provide essential cooling for up to 8x GPUs with four heavy-duty fans in front and four Redundant 3000W Titanium level PSUs in back with their own integrated fans pulling out heated air. Power is provided by four redundant 2200W Titanium level PSUs.

There are actually two configurations of this GPU server, the Supermicro GPU SuperServer 420GP-TNAR+, and the Supermicro GPU SuperServer 420GP-TNAR. The only difference between the two is the power supplies. The 420GP-TNAR+ offers 4x 2+2 redundant 3000W Titanium level power supply modules, and the 420GP-TNAR has 4x 3+1 redundant 2200W Platinum level Power Supply modules.

*This system is only sold as a completely assembled system with a minimum of 2x CPUs, a minimum of 1TB of memory (2TB highly recommended) for 80GB HGX-8 A-100 OR a minimum of 512GB for 40GB HGX-8 A100, 1x storage device and 1x NIC

Performance

With dual Socket P+ (LGA-4189) CPU sockets, this system can support 2x 3rd generation Intel Xeon Scalable processors. Using top of the line Platinum processors, it can support up to 40 physical cores and 80 virtual threads. Gold series processors will provide up to 32 cores maximum and up to 20 cores using Silver Xeon Scalable processors. All feature PCIe 4.0 for up to twice the throughput compared to PCIe 3.0. There are also 33% more PCIe lanes at 64x per CPU compared to only 48x on the previous generation. Memory capacity has also been improved with 8x memory channels compared with only 6x before, not to mention an increase in supported memory from 2TB to up to 4TB per CPU.

Memory

The Supermicro GPU SuperServer 420GP-TNAR+ has 16x memory module slots per CPU. With both processors installed it has 32x active memory module slots. Each memory channel can host two memory modules for maximum density and a maximum capacity of up to 12TB. 12TB can only be achieved using data centric persistent memory modules (DCPMM), commonly referred to as Intel Optane memory modules. At maximum capacity 16x memory module slots are populated with DCPMMs and the other 16x using Registered (RDIMM) memory modules. Both RDIMMs and Load Reduced (LRDIMM) memory modules will provide up to 8TB of memory at capacity.

Storage

In total 10x 2.5-inch hot-swap drive bays are supported on this system. 6x are located on the front upper portion of the chassis. The other four are optional and located in the rear of the server above the 4x large PSUs. The front bays support SATA, SAS or NVMe drive options, while the optional 4x rear bays are dedicated for NVMe drive formats. 2x PCIe 4.0 x4 slots on the system board can be outfitted with NVMe or SATA3 M.2 drives to support the operating system only. If installing SAS drives, then an optional SAS RAID controller is required and will provide data transfer speeds of 12Gb/s compared to the onboard SATA connection at 6Gb/s. Drive trays are tool-less for easy installation and removal. While the drives are hot-swap, orderly hot-swap is recommended for the NVMe drives.

Expansion

This system is designed to support the HGX A100 8-GPU 40GB/80GB SXM4 multi-GPU board for unprecedented acceleration at every scale compressing the time-to-solution down dramatically. GPU-to-GPU communications is handled by an Nvidia NVlink with NVSwitch interconnect allowing the GPUs to communicate with each other at speeds of up to 600GB/s. The communications on the Nvidia NVlink with NVSwitch interconnect are 10x faster than the fastest PCIe Gen4 x16 bus enabling lightning-fast peer to peer communications. Communications from the GPU to CPU are still handled by a PCIe Gen4 x16 bus. 8x PCIe x16 slots at the back of the system on the switch tray support low-profile slots from the PLX PCIe expander modules with half allocated to CPU 1 and the other half to CPU 2. These slots are used for up to 8x high-performance I/O cards, like InfiniBand with 100Gb/s or 200Gb/s access speeds and direct communication with the GPUs. With 8x GPUs and 8x NICs for GPU direct memory access there is a 1:1 ratio of GPUs to remote direct memory access (RDMA) capable NICs. Two additional PCIe 4.0 x16 slots on the front of the system are split between CPUs 1 and CPU 2. An open compute project 3.0 (OCP 3.0) advanced I/O module (AIOM) can be installed in the Mezzanine slot at the back of the server on the far left saving your limited PCIe slots for other options.

Management

On the left side front of the system there is a dedicated RJ45 LAN port to the Baseboard management controller (BMC), plus a few tell-tale status lights for system health. Status lights include NIC (LAN) LEDs, General information LED, Power LED, and UID LED. More LED lights on the back of the system indicate switchboard activity with 12x LED lights that indicate CPU to PLXs and PLXs to GPUs. The PSUs and storage trays also feature LEDs for at a glance status. This system can also be used with Supermicro’s proprietary Watch Dog and SuperDoctor5 applications for remote and at chassis management of the system. Integrated platform management interface 2.0 (IPMI 2.0) and Redfish API are also supported for interaction with other third-party hardware.

Summary

Outfitted with up to 8x high-performance A-100 Ampere-based Nvidia SXM4 GPUs, plus up to 8x high-performance I/O devices, this system offers superior credentials for high-performance computing applications. Intel’s 3rd generation Xeon Scalable processors also deliver significant memory density plus support for PCIe 4.0 enabling superior throughput compared to PCIe 3.0. The Supermicro GPU SuperServer 420GP-TNAR+ features 5 petaFLOPS in a 4U chassis, making it a perfect platform for any AI workload including training, inference, and analytics.