While interest in machine learning algorithms to drive artificial intelligence (AI) applications is high, a survey published this week suggests many organizations are finding it challenging to move past the AI development phase.
The survey of 344 technology and IT professionals conducted by Dimensional Research on behalf of Univa, a provider of workload management software, finds that while 93 percent of respondents have launched a diverse range of machine learning projects, only 22 percent of those projects have been deployed in a production environment. The top reason cited for not making that transition is the myriad technical challenges associated with migrating AI applications into a production environment, the report finds.
Building and training AI models based on machine and deep learning algorithms is a laborious task that requires access to massive amounts of data. AI models that depend on deep learning algorithms can also be especially expensive to build and deploy because they generally require graphic processor units (GPUs) to run efficiently. More challenging still, AI models in a production environment require access to inference engines that need to be optimally deployed within or alongside a business application to optimize a business process. The truth of the matter is that AI applications in production will require significant upgrades to IT infrastructure on-premises and in the cloud.
Univa CEO and president Gary Tyreman says the survey makes it clear that given the complexity and compute resources required to run these AI models, there’s a high correlation between organizations investing in AI that also have access to high-performance computing (HPC) platforms on-premises or in the cloud. More than 88 percent of respondents indicate that they are working with HPC in their jobs, which suggests the cost of entry for developing AI models is significant.https://o1.qnsr.com/log/p.gif?;n=203;c=204663295;s=11915;x=7936;f=201904081034270;u=j;z=TIMESTAMP;a=20410779;e=i
“There’s a direct correlation between HPC and machine learning,” says Tyreman.
In fact, the survey finds nearly 90 percent of respondents expect to employ GPUs, while more than 80 percent of respondents plan to use a hybrid cloud to try to contain costs.
While there is a general expectation that AI will become pervasive, the Univa survey makes it clear that most organizations are still wrestling with fundamentals. On top of that, there is already a widespread shortage of the data science skills required to build AI models. Given all those costs and immature processes associated with building and deploying AI models, all the hype concerning AI applications might be a tad overblown. At this point, it’s starting to become a lot more apparent that any future dominated by widespread usage of AI is much farther away than most people outside of AI might otherwise be led to believe.