Author

Chetan Prakash

Director, System & Technology Group

IT infrastructure isn’t the star of today’s buzzworthy AI projects, but it’s a foundation no business can succeed without. In this post I’m discussing why infrastructure is important and looking at good options for deploying – or evolving – infrastructure to become AI-ready.

When I think of artificial intelligence (AI) systems, I think of two things. First, ultra-powerful servers with fast GPUs and CPUs working in parallel. And second, AI software frameworks that make incredible things possible – such as turning shape-drawings of people into photo-realistic images. These are the AI technologies that excite people.

However, I also know they cannot perform without the right infrastructure to support them. Understanding this is one of the keys to AI success.

In the Gulf region, there’s no shortage of enthusiasm for AI. Governments have invested in AI initiatives, while surveys show that citizens are ready for AI-based services. Yet actual adoption of AI has been slower than expected, partly due to a lack of understanding of the technology. “Many organizations don’t know where to start,” wrote CIO.com last month.

AI is delivering real-world value for businesses around the world, including some in the Gulf. For everyone else, now is a good time to catch up – starting with infrastructure.

Why is AI-ready infrastructure important?

The reason infrastructure is important to AI isn’t complicated, especially when you remember how much data gets thrown around by AI workloads.

Analyzing and manipulating big datasets is at the heart of deep learning and machine learning. AI workloads are organized in a “pipeline” of tasks, and data must be delivered responsively at each stage. Slow data delivery to one task can hold up the entire workload.

AI workloads also tend to be made up lots of small calculations, rather than a few big and complex calculations. This allows workloads to be distributed across systems, and accelerated by multiple processors (like GPUs) running in parallel. But for this to be effective, each system and processor needs to be able to share data with extreme speed. Otherwise, again, the pipeline will be blocked and results will be slowed.

So imagine your AI team has developed an exciting project, but they have to run it with…

• Standard servers with no acceleration for AI workloads
• Traditional HDD storage that reads and writes data really slowly

Those are two big bottlenecks that will slow down results and prevent you from finding the insights you wanted. Furthermore, the latest AI framework versions might not support your infrastructure.

AI-ready infrastructure is designed to solve these problems.

Don’t throw out old servers and storage yet!

The good news is, businesses don’t need to “rip and replace” existing infrastructure to start the AI journey. You can add the capability you need to run your first AI projects, then keep evolving as needed.

GBM is a great partner if you want to evolve in the smartest way. Our solutions are designed around business needs – so we’ll consult with you about your goals and your current infrastructure, and we can help you deploy the best hardware from a range of vendors.

It’s worth considering servers and storage from IBM, which is the worldwide AI market leader according to IDC research.

Servers designed for the AI journey

IBM’s latest Power System servers are part of its Enterprise AI line – which means they’re designed to help businesses succeed at AI even if there are no data scientists in the company.

Some of the IBM Power System features that answer the problems we discussed above include:

• Fast PCI Express I/O for sharing data across the system
• Up to 120TB onboard storage for rapid reading and writing
• IBM POWER9 processors, which are designed specifically for AI workloads and feature high-bandwidth CPU-GPU interconnects

The latest AI software, such as H2O.ai machine learning tools, are specially optimized for Power Systems. IBM also provides its own AI software like PowerAI Vision, which can be trained to classify images and detect objects in images and videos.

Supercomputer storage for your AI projects

IBM is also one of few companies to offer ultra-fast storage solutions designed specifically for AI.

Spectrum Storage powers the world’s two most powerful supercomputers, Sierra and Summit, but it’s not just for supercomputers. It can help any business to build the right information architecture (IA) to support AI success – because there is “no AI without IA”. This means…

• Unifying data access to make data collection for AI projects less time consuming
• Accelerating data throughput to support GPU acceleration and better AI model accuracy
• Support for containerized workloads, as is the trend with AI projects that need to be quickly deployed, scalable and portable

In one genomics case study, IBM storage shortened runtime by 96%, cut costs by two thirds and helped project leaders design and deploy in just two weeks.

Let’s talk AI infrastructure…

So, as you can see, infrastructure matters when it comes to AI projects.

If you’d like to know more – including how you can evolve your existing servers and storage to tackle AI, without over-sizing and over-spending – please contact me.