What does an AI Exascale Platform look like?

What does an AI Exascale Platform look like?

We have all heard quite a lot about AI (artificial intelligence), according to many experts, including the likes of Bill Gates of Microsoft, Artificial intelligence is a massive, structural game-changer on par with the internet or computing. It could reshape global productivity, automates routine tasks, and accelerates workflows across nearly every industry. “AI is cognitive intelligence, not just another tool“. GPUs (Graphics Processing Units) are used in AI to perform thousands of mathematical calculations at the same time. While a standard computer processor (CPU) handles tasks one by one, a GPU uses parallel processing to run massive numbers of repetitive matrix and vector multiplications simultaneously, cutting the time needed to train and run AI models from months to hours

• • LLM large language models: Is a type of artificial intelligence program trained on massive amounts of text to understand, summarize, translate, and generate human-like language.
  • ChatGPT for example: runs on massive clusters of high-performance data centre accelerators built by NVIDIA. OpenAI initially trained and ran early versions of ChatGPT using tens of thousands of Nvidia A100 and the then newer NVIDIA H100. Today we have next generation B200 and B300 NVIDIA GPUs.
  • Productivity boost: Automates repetitive tasks like drafting code, summarising data, and routine communication.
  • Software and digital creation, AI drastically cuts production timelines. Development cycles that used to take hundreds of hours are seeing massive time reductions as small teams achieve the output capacity of much larger historical workforces
  • Augmentation over replacement: Functions best when handling 70% of preparatory or grunt work, leaving 30% for human judgment.

So What does an AI Exascale Platform look like? Well as a little more background. The world’s wealthiest companies NVIDIA, Alphabet (Google), Apple, Microsoft, and Amazon hold market values exceeding $2.5 trillion each, driven by traditional computing and software and more recently, in the case of NVIDIA, largely by artificial intelligence. NVIDIA produces the GPUs (graphic processing units) used for AI computing. The current leading GPU is the B300 which is the building block of a Exascale platform. The following example from Supermicro has 72 NVIDIA Blackwell GPUs and 36 Grace CPUs integrated into this single rack with a 1.8TBb backbone and each rack requires up to 200kW of power. An exascale system is a supercomputer capable of performing at least one exaFLOP, which equals one quintillion (10¹⁸) floating-point operations per second. This massive processing milestone is one million times faster than the terascale computers built in the early 2000s. Now we here at Datanet have been in the rack business a long time and, by way of comparison, a typical compute and network rack uses around 4kW. Up to 16 of these Exascale compute platform can be connected together in a East/West fashion taking you on the journey to Zettascale performance!

So what does an AI Exascale Platform cost?
This GB300 NVL72 Exascale Platform sells for about £6M with a running cost, just for power of about £60k per month. It comprises a 47U rack, 18 compute trays each with 4 Blackwell B300 GPUs and 2 Gracewell CPUs, a total of 72 B300 GPUs and 36 Gracewell CPUs, 9 switch trays providing 1.8TB interconnect and a CDU cooling unit.

Your workload (and budget) may require a powerful yet sub-exascale level of AI GPU computing then take a look at the following systems which allow you to run AI inference on the class leading 8U chassis with 8 x B300 GPU system which is available in a matter of weeks for a much lower budget ~£550k, or the entry level 2U chassis AI RTX GPU workstation, both shown in the following images. Which to choose, GB300, B300 or B200 depends on the workload and Supermicro can help here too with their free to use benchmark and validation tool JumpStart.

Here at Datanet we can help with your Exascale or other AI and HPC requirements from sourcing to hosting and maintaining

🔹 Ready for Action: Our turnkey AI and high-performance racks with high power and cooling are available in our secure government grade data centre, to come and have a look, just fill in the contact form below!!

🔹 Power and Cooling: AI GPU racks can support up to 200kW together with DLC (direct liquid cooling) from the adjacent CDU (coolant distribution unit) cooling rack.

🔹 Accelerate Your Vision: We also have air cooled 47U racks supporting up to 25kW per rack alongside regular 2kW to 7kW racks, and smaller one-third (⅓) 15U racks too, our infrastructure is your gateway to innovation.

Whichever AI Inferencing model you choose you can trust this 30 year British company Datanet with your Data Sovereignty?

With all the uncertainty in today’s world, not least from our long-standing US allies, UK business needs to retain some independence at the IT and data storage level. Businesses in the Cloud are thinking again about their Cloud First policy and re-considering the sovereignty question. We hear this a lot now at Datanet, businesses considering moving back to on-prem or better yet their local data centre.

Datanet have been providing UK sovereign based Data Centre and Hosting services for 30 years, furthermore with our helpful, adaptable on-site support staff, now is the time to re-consider your London Edge DC for your AI and Infrastructure requirements.

Datanet Ultra Data Centre (PDF)

ARK Sustainability (PDF)

To find out more about how Datanet can help, call 01252 810010, or email Datanet Sales or simply fill out the form on our contact page and we will be in