Smart Cities

Customisation is brought to AI, machine learning and high performance computing (HPC) applications with the Universal GPU server system, said Super Micro Computer. According to the company the “revolutionary technology” simplifies large scale GPU deployments and is a futureproof design that supports yet to be announced technologies. 

The Universal GPU server combines technologies supporting multiple GPU form factors, CPU choices, storage and networking options optimised to deliver configured and highly scalable systems for specific AI, machine learning and HPC with the thermal headroom for the next generation of CPUs and GPUs, said the company.

Initially, the Universal GPU platform will support systems that contain the third generation AMD EPYC 7003 processors with either the MI250 GPUs or the Nvidia A100 Tensor Core 4-GPU, and the third generation Intel Xeon Scalable processors with built-in AI accelerators and the Nvidia A100 Tensor Core 4-GPU. These systems are designed with an improved thermal capacity to accommodate up to 700W GPUs.

The Supermicro Universal GPU platform is designed to work with a wide range of GPUs based on an open standards design. By adhering to an agreed upon set of hardware design standards, such as Universal Baseboard (UBB) and OCP Accelerator Modules (OAM), as well as PCI-E and platform-specific interfaces, IT administrators can choose the GPU architecture best suited for their HPC or AI workloads. In addition to meeting requirements, this will simplify the installation, testing, production, and upgrades of GPUs, said Super Micro Computer. In addition, IT administrators will be able to choose the right combination of CPUs and GPUs to create an optimal system based on the needs of their users.

The 4U or 5U Universal GPU server will be available for accelerators that use the UBB standard, as well as PCI-E 4.0, and soon PCI-E 5.0. In addition, 32 DIMM slots and a range of storage and networking options are available, which can also be connected using the PCI-E standard. 

The Supermicro Universal GPU server can accommodate GPUs using baseboards in the SXM or OAM form factors that use high speed GPU-to-GPU interconnects such as Nvidia NVLink or the AMD xGMI Infinity fabric, or which directly connect GPUs via a PCI-E slot. All major current CPU and GPU platforms will be supported, confirmed the company.

The server is designed for maximum airflow and accommodates current and future CPUs and GPUs where the highest TDP (thermal dynamic performance) CPUs and GPUs are required for maximum application performance. Liquid cooling options (direct to chip) are available for the Supermicro Universal GPU server as CPUs and GPUs require increased cooling. 

The modular design means specific subsystems of the server can be replaced or upgraded, extending the service life of the overall system and reducing the e-waste generated by complete replacement with every new CPU or GPU technology generation.

http://www.supermicro.com

At GTC 2022, Nvidia announced its next generation accelerated computing platform “to power the next wave of AI data centres”. The Hopper architecture succeeds the Ampere architecture, launched in 2020.

The company also announced its first Hopper-based GPU, the NVIDIA H100, equipped with 80 billion transistors. Described as the world’s largest and most powerful accelerator, the H100 has a Transformer Engine and a scalable NVLink interconnect for advancing gigantic AI language models, deep recommender systems, genomics and complex digital twins. 

The Nvidia H100 GPU features major advances to accelerate AI, HPC, memory bandwidth, interconnect and communication, including nearly 5Tbytes per second of external connectivity. H100 is the first GPU to support PCIe Gen5 and the first to utilise HBM3, enabling 3Tbytes per second of memory bandwidth, claimed Nvidia. 

According to Jensen Huang, CEO, 20 H100 GPUs can sustain the equivalent of the entire world’s internet traffic, making it possible for customers to deliver advanced recommender systems and large language models running inference on data in real time. 

The Transformer Engine is built to speed up these networks as much as six times compared with the previous generation without losing accuracy. 

MIG technology allows a single GPU to be partitioned into seven smaller, fully isolated instances to handle different types of jobs. The Hopper architecture extends MIG capabilities by up to a factor of seven over the previous generation by offering secure multi-tenant configurations in cloud environments across each GPU instance.

According to the company, H100 is the first accelerator with confidential computing capabilities to protect AI models and customer data while they are being processed. Customers can also apply confidential computing to federated learning for privacy-sensitive industries like healthcare and financial services, as well as on shared cloud infrastructures. 

To accelerate the largest AI models, NVLink combines with an external NVLink Switch to extend NVLink as a scale-up network beyond the server, connecting up to 256 H100 GPUs at nine times higher bandwidth versus the previous generation using NVIDIA HDR Quantum InfiniBand. 

New DPX instructions accelerate dynamic programming by up to 40 times compared with CPUs and up to a factor of seven compared with previous generation GPUs, said Nvidia. This includes, for example, the Floyd-Warshall algorithm to find optimal routes for autonomous robot fleets and the Smith-Waterman algorithm used in sequence alignment for DNA and protein classification and folding. Availability NVIDIA H100 will be available starting in Q3. 

https://www.nvidia.com