Today our friends at NVIDIA announced that Verne Global’s Icelandic data center has been selected as one of its initial three European DGX-Ready Data Center program partners. We’re delighted to be working with NVIDIA on this program and to have our data center identified as an optimal location for their powerful range of DGX AI supercomputers.
In the race to develop and scale AI applications, organisations and AI entrepreneurs shouldn’t be limited by data center capabilities or the underlying energy and utility infrastructure supporting accelerated compute platforms. This is why Verne Global and our Keflavik data center is such a perfect venue for hosting the NVIDIA DGX heavy metal. However, none of this is by chance – everything that makes our campus the location of choice for industrial- scale computing has been built from below-the-ground up.
Anyhow, before I get all Icelandic on you, here’s some context to why I became interested in NVIDIA’s DGX ambitions.
I saw my first DGX-2 at GTC in San Jose in March 2018. It was an awesome sight standing twice as high as a regular large server and looking like it originated from one of NASA’s space development labs.
Its 2 petaFLOPS performance (delivered in a single system) was even more stunning, resulting from all of its 16 NVIDIA V100 GPUs tightly connected via NVIDIA NVLink and NVSwitch plus Mellanox InfiniBand networking providing the ability to scale by connecting multiple DGX-2 systems together. There was now an AI training solution that could handle the largest and most compute intensive deep neural network models.
Designed to solve the world’s most complex AI challenges, the DGX systems have been rapidly adopted by a wide range of organisations across dozens of countries globally. Its fierce reputation quickly grew such that no self-respecting FTSE100 AI department could be without one and Facebook was one of its first announced customers. It’s found a receptive following in the machine vision, autonomous vehicle system development and drug discovery domains, where large models and billion record datasets are normal.
We’ve all grown up with NVIDIA primarily selling its semiconductor chips to OEMs to build into products like graphics cards for desktop gaming PCs. NVIDIA recognised the need to lead the industry with a new standard in how AI computing should be systemised for enterprise - an architectural approach which is now being followed by all the major server OEMs.
As with any leading-edge product, DGX places unique demands on data center resources, which must be optimised for AI-accelerated workloads. As such, there are nuances in the ways it should be configured for peak performance and these have nothing to do with the systems themselves but with the architectural requirements of the data center facilities that host them - including racking, power, cooling and networking. As a result, more thoughtful consideration and planning is required versus dropping a DGX-2 into your traditional downtown, enterprise data center.
Traditional enterprise Raised-floor Low Density Environment (3-5kW/rack)
NVIDIA immediately mobilised its resources to focus on the systems deployment challenges facing many enterprises, and they rapidly developed the DGX-Ready Data Center program to help customers avoid the constraints of hosting AI infrastructure. Many of the vendors and partners I speak to on a daily basis admire this program and a fair few of them wish they had had the foresight to launch a similar one for their intensive hardware.
The road to becoming certified has been tough and easy in equal measure. Let’s start with the tough. So, what was apparent from the start was this isn’t a piecemeal certification program heavy on marketing fluff and light on content. I’ve been constantly impressed at the level and depth NVIDIA’s team have gone to ensure a data center location is perfectly able to cater for their machines.
NVIDIA’s technical team has gone through 100s of pages of our commissioning reports, including stacks of data and test results on the power and cooling performance of the data center and our infrastructure. I am sure they almost wished they had a DGX-2 available to help us crunch the numbers! It was also news to me - mostly a computer hardware type - that the data center halls are tested with up to 150% of their rated power load using dummy equipment and every parameter monitored.
But that’s not all. There was also a technical audit meeting, where the previous challenges are fielded to determine how our data center technical team would resolve them. Add to this the 90-minute grilling of Tate Cantrell, our CTO, covering topics like extreme cooling densities and hot air handling, high-density power distribution, flooring structural integrity (DGX-2 weighs 360lbs/163kg), support capabilities, etc. But, after all the stages and reports I am pleased to say our infrastructure, technologies, support procedures and our world-class Icelandic operations team passed with flying colours.
So now the easy part, and by easy I’m talking about what we at Verne Global do best on a daily basis. It was easy to demonstrate our effectiveness at hosting accelerated computing operations because in simple terms, it’s our bread and butter. Since 2012 we have built for industrial scale and we specifically cater for high performance computing (HPC). Our Icelandic location was originally chosen because we wanted access to vast amounts of low-cost, dual sourced, 100% renewable and scalable energy, all provided by an industrial power grid renowned for its reliability. Add on top of that our campus, which is based on a former NATO military base, was selected for its security and network connectivity. And, built on that is infrastructure designed by our CTO, Tate, that is purpose built for high-density, high-capacity equipment, all cooled by 100% ambient Icelandic free air.
So, for companies that don’t have data center facilities that can support advanced computing hardware such as DGX supercomputers, Verne Global is ready to deliver on that challenge while freeing you to focus on insights and innovation thanks to NVIDIA’s DGX-1 and DGX-2 – together, the world’s most powerful AI platform.
Let’s chat and find you the best possible home for your DGX-1 and DGX-2 systems.
Bob Fletcher, VP Strategy, Verne Global (Email: [email protected])