Artificial Intelligence is making huge strides in the automation of tasks requiring image analytics, with image processing GPUs providing the necessary horsepower. Glenn Fitzgerald, Chief Technology Officer for the Product business in EMEIA at Fujitsu, talks about Fujitsu’s participation at leading GPU vendor NVIDIA’s annual European showcase.
Have you ever wondered what it takes to paint a Picasso? That’s one of the applications of Artificial Intelligence (AI) that we demonstrated at the recent NVIDIA GPU Technology Conference (GTC) in Munich, alongside a panoply of other AI and Virtual Reality solutions. Passers-by were amazed at the levels to which the technology was able to replicate the image – because of its deep attention to detail.
It’s not that we are encouraging art forgery, of course. Nor are we suggesting that AI has yet reached the point where it competes with human artistic creativity – although some believe that is not far away. The key point we wanted to get across was that High-Performance Computing (HPC) architectures or systems and algorithms have matured over the years to now have the capability to create an AI solution suitable for business use cases. Deep learning systems – readily available as pre-integrated solutions or reference architectures – can now recognize, deconstruct, manipulate and reassemble images with a degree of sophistication that matches human abilities. And this opens up any number of possible applications.
Fujitsu at NVIDIA GPU Technology Conference
At GTC we demonstrated two solutions: a ready-to-go, integrated deep learning solution based on Fujitsu’s Zinrai Deep Learning System and one based on Fujitsu’s reference architecture for quality control in manufacturing, set up with PRIMEFLEX and Fujitsu’s advanced image recognition (FAIR) software stack. In other words, this isn’t sci-fi: it’s something that many organizations already have or can easily access in pre-integrated systems.
Let’s leave art on one side for another discussion. At GTC, we were concentrating on the use of AI and image analytics for industrial and commercial purposes. Here, Fujitsu is a leading player, with solutions including quality control, fraud detection, security improvement, transport optimization and marketing personalization.
Our AI heritage
Fujitsu has one of the IT world’s longest, most successful heritages in AI research and development. In the 1980s we built Japan’s first computer equipped with AI, named FACOMα, and we have now filed more than 200 AI-related patent applications – for example, maximizing the computational horsepower available to neural networks using supercomputers – making us the leader among Japanese IT vendors.
The reason for focusing on image analytics at GTC is that AI holds the potential to automate entire new classes of activity that previously demanded the visual skills of humans – recognizing features such as faces, for example. But image processing is hugely resource intensive and NVIDIA is the acknowledged leader in GPU technology, with the necessary power in its hardware to cope with the task. Putting Fujitsu and NVIDIA technology and expertise together creates a compelling combination.
Our partnership has been in place for decades already and in March this year, we announced its latest fruit – the availability of Fujitsu’s ultra-high performance CELSIUS workstations and PRIMERGY server solutions, harnessing the latest visual computing platforms from NVIDIA with support for the world’s most advanced visual computing platform, NVIDIA Quadro GV100. This creates the world’s most powerful architecture for high-performance computing, AI, virtual reality, simulations and graphics workloads on professional desktops, targeted at sectors including healthcare and life sciences, automotive, financial services, science, and manufacturing.
Zinrai Deep Learning System
On show at GTC was the Fujitsu AI Solution Zinrai Deep Learning System. Using this system, customers looking to build deep learning capability in an on-premises or hybrid environment can quickly set up a platform offering world-class speed and the option to perform learning processing with low usage frequency in the cloud and/or learning processing with a high usage frequency in an on premises environment. The deep learning server employs the latest NVIDIA Tesla® P100 GPU, with up to eight GPUs in a single server, giving customers the flexibility to select a system structure based on application.
Quality control in manufacturing
Also at GTC we showed an example of Fujitsu’s reference architectures for quality control (QC) in manufacturing, based on PRIMEFLEX and Fujitsu’s Advanced Image recognition Software Stack. In this case we were using Fujitsu’s PRIMEFLEX for HPC with Fujitsu’s Advanced Image Recognition Software (FAIR) for AI quality control in production lines, powered by Fujitsu PRIMERGY CX400M4 (an Intel Select Solutions platform). This is an integrated deep learning solution for improved defect localization and classification, ultimately leading to increased yield, lower risk, and enhanced flexibility in manufacturing product lines.
The demo was based on a factory floor environment, where advanced image recognition technology can pick up even tiny product defects, dramatically reducing the time needed for quality control checks by humans. In addition, we showed the image recognition analysis being delivered at the edge– effectively closest to the line itself – illustrating how businesses can transform their production processes with automated real-time decisions.
The image recognition core supports process automation and orchestration for workload optimization in multi-node clusters. And the overall solution leverages seamless use of parallelism and hardware acceleration for scale – in other words these are core HPC architectures and hence customer can either leverage their own HPC architectures or choose reference architectures – such as our demo – from Fujitsu.
To give you a flavor of how the rolls out in real life, Fujitsu has co-created an AI quality control solution with wind turbine manufacturer Siemens Gamesa that performs microscopic-level quality controls. Manually evaluating Ultrasonic Testing (UT) scans of each blade used to take up to six hours. The AI solution from Fujitsu and Siemens Gamesa can automatically detect flaws, achieving 100 percent coverage of all defects, reducing nondestructive test scan times by 80 percent – and, here’s the twist, also identifying miniscule faults invisible to the human eye.
I’m not entirely sure what Picasso would have made of all this, but where image processing is concerned, it’s clear we have moved from the avant garde stage of AI to the point where any large enterprise can – should – consider how this technology might transform important areas of manufacturing or business processes. In many cases this might involve your existing HPC hardware. In many cases this might involve your existing HPC hardware.