Earlier this month, we reported Exactrous examinationA group of 18 machines operating 144 NVIDIA H200 GPUS, which happens to be one of the first clusters according to these processors. Since then, Hydra Host, the company which facilitated the construction of the cluster, has given us additional details on the system. The cluster uses Lenovo systems with several personalizations of Hydra Host, which played an important role. The machine can also be rented – when it is not used by the owner – via the Brokkr platform of Hydra.
A lot of calculation power
The dorsal thorn of the cluster consists of 18 Lenovo knots equipped with 144 GPU NVIDIA H200 and 20 TB of HBM3E – or eight by system – allowing calculation performance of 570 PETATOPS FP8 for AI. 16 knots are configured and refined by Hydrahost for training, which requires massive calculation and memory performance, while the other two serves as an inference nodes. In addition, Hydra Host installed its Brokkr platform for supply, management and remote rental (more remotely (more on this subject later).
Hydra Host collaborated with Computacenter to design a high performance networking architecture adapted to the needs of the cluster. The configuration uses the infiniband of 3.2 tops for the East-West traffic and 400 GB Apron Ethernet switches. Computacenter network engineers have provided all the components aligned with NVIDIA’s reference architecture for transparent compatibility.
“We provided the 18 Lenovo nodes with H200 GPUs (16 interconnected knots and two inference nodes), designed the networking architecture in collaboration with Computacenter, and facilitated roommate via Patmos,” said Andrea Holt, Word of HOST HOST.
The cluster itself is fairly powerful, even in terms of computers for general use. The servers present 192 96 core processors (for a total of 3,456 cores) associated with 36 TB of DDR5 memory and 270 TB of storage in the solid state NVME. There are spare berries so that the storage space can be widened easily. The supercomputer uses a custom built network by Hydrahost.
The company also called on Patmos to manage the roommate, offering enough energy (approximately 100 kW) and cooling for power and hot machines.
Best performance at the best price
The accuracy costs $ 5 million, with an average of $ 277,777 per machine, comparable to a single H200 plinth to 8 lanes rather than a full server. Here is where it becomes interesting. Who made this price facilitated?
On the one hand, Hydra Host is a near Nvidia partner and only offers NVIDIA GPUs as a service. In addition, its Brokkr software is mainly optimized for Cuda. On the other hand, Exami is a company supported by Nvidia, so it can potentially obtain preferential prices.
“We are the best on the market to get our customers the good GPU for their needs and at the best price,” said Ryan Horjus, main sales engineer at Hydra. “This cluster was supported by Nvidia from an architecture design and their creation program. Hydra managed it for exams, as we do for other companies.”
Hydra is also specialized in the construction of personalized solutions for startups and even monetizes their machines when they are not used.
“Hydra has helped startups enter their own clusters for better prices thanks to bulk purchases,” added Horjus. “They can obtain ideal prices via our network. They are also able to monetize servers when they are not used via the Brokkr management platform.”
Speaking of Brokkr, this is GPU management and supply software and a monetization platform for GPUs. It provides data centers and startups with a turnkey software solution to put their equipment between customers and pay them, said Ariel Deschapell, technology director and hydra co -founder.
“One of its main characteristics is the automated supply of bare metals and life cycle management,” described Deschapell. “This means that the platform does all the configuration and management work of the operating system and the firmware of the basic server, configuration of drivers and other support software, and perform tests on GPU and other components.