Furthering the long-standing partnership for next-gen GPU computing solutions

To address the multitude of use cases, GIGABYTE offers the largest portfolio of GPU computing server solutions in the market, with modular system design and configurability in mind.

The solutions come with optimised air cooling and maximum GPU power per rack unit space. The portfolio continues to expand as next-generation computing technologies from major CPU/GPU manufacturers enter the market all aiming for the highest computing density, performance, and energy efficiency.

Among the various NVIDIA-Certified systems, the following GPU servers from GIGABYTE are of particular interest for this press release announcement: G293-Z22 and G593-ZD2.

GIGABYTE G293-Z22 – the densest GPU computing platform in 2U form factor

Based on the latest AMD EPYC 9004 CPU architecture, the G293-Z22 system design uses a single CPU socket to control up to 8 x NVIDIA GPU cards (PCIe form factor, Gen5 x16, double- or single-slot cards) by leveraging a high core count AMD EPYC 9004 CPU (up to 84 cores with up to 300W TDP CPU).

The unified memory space (as in a single NUMA) across CPU, GPU, system memory, storage devices, and network devices gives the greatest computing performance with the least latency in data movement. Either in bare metal set-up or in virtualisation, G293-Z22 can guarantee optimal distribution of computing resources.

G293-Z22 comes with 8 x PCIe Gen5 slots for NVIDIA GPUs, 1 x CPU socket for AMD EPYC 9004 CPU, 12 x DDR5 4800MHz DIMM slots, 8 x 2.5” hot-swap drive bays (of which four bays support PCIe Gen5/SATA/SAS protocols and the other four bays support SATA/SAS), 2 x M.2 storage slots (NVMe PCIe Gen5 x4), 2 x PCIe Gen5 expansion slots (x16 bus width, low-profile form factor) for add-on devices such as HBA FC / storage cards, NVIDIA ConnectX-7, or NVIDIA BlueField-3 DPU to accelerate data transfer across nodes and clusters and GPUDirect/RDMA. Such compact, GPU-centric computing features interest HPC users especially who work with AI, molecular simulations, genomics sequencing, weather prediction, and other use cases.

GIGABYTE G593-ZD2 – the most powerful GPU system with NVIDIA H100 SXM5 & NVLink Fabric

G593-ZD2 is among the best seller models at GIGABYTE: the system is based on the NVIDIA H100 SXM5 8-GPU platform, supports two AMD EPYC 9004 CPU sockets, and offers the possibility of installing up to 12 x NVIDIA SmartNICs or DPUs to accelerate data transfer across nodes and clusters and GPUDirect/RDMA. G593-ZD2 is well suited also for providing maximum Multi-Instance GPU (MIG) sessions for AI developers who run workloads under different containerised environments and require custom algorithms, libraries, and datasets to be executed in isolated user spaces.

The system employs a novel cooling solution that dedicates a cooling chamber for NVIDIA GPUs and SmartNIC/BlueField DPU used in the PCIe expansion slots, ensuring the highest airflow possible to cool the high-performance components. In fact, the system consists of two separate parts: a 1U CPU server that sits above a 4U GPU tray. The 1U tray houses the CPU, system memory, storage bays, and rear facing PCIe slots. The 4U GPU tray is easy to slide-out in the event of system maintenance, considering the intricate onboard interconnects that link all the GPU modules and the 1U server together. In addition, the onboard 12 x PCIe Gen5 expansion slots (8 x low-profile and 4 x full-height) are housed in toolless trays to facilitate maintenance and replacement of parts. Onboard storage includes 8 x 2.5” hot-swap drive bays (NVMe PCIe Gen5) and 2 x M.2 storage slots (NVMe PCIe Gen3). The system memory capacity includes 24 x DDR5 DIMM slots at 4800MHz.

The inclusion and choices of the NVIDIA H100 SXM5 modules in the G593-ZD2 system is important, in that new NVIDIA Magnum IO GPUDirect technologies favour faster throughput while offloading workloads from the CPU to achieve performance boosts. G593-ZD2 supports NVIDIA GPUDirect RDMA for direct data exchange between GPUs and third-party devices such as NICs or storage adapters. And there is support for GPUDirect Storage for a direct data path to move data from storage to GPU memory while offloading the CPU, thus resulting in higher bandwidth and lower latency.

Conclusion

Beyond the current HPC technology and onward to Q3/Q4 2023 and further, GIGABYTE is ready to timely launch next-generation GPU computing and hybrid computing solutions in partnership with NVIDIA. GIGABYTE will continue to address diverse use cases by adapting system design to real-world workflows and data centre architectures.