Will Your IT Infrastructure Have Sufficient Power for the AI Revolution?

Artificial intelligence (AI) has come a long way in the past five short years. Its power hungry technology is creeping up on us and could catch data center managers unaware unless we anticipate the effects. This short paper examines what artificial intelligence is becoming.  Then we move on to the GPU, ASIC and FGPA hardware components that enable AI to compute.

The Current State of AI in Computer Science

Having machines that think and react like humans is a critical component of the rising Fifth Industrial Revolution. Exponentially increasing computer power is enabling new speech recognition, learning, planning, and problem solving technologies.

However, these new abilities require more power than available in conventional computers. They need access to objects, categories, properties, and relations between all of them to implement knowledge engineering.

Initiating common sense, reasoning and problem-solving power in machines is a difficult and tedious task. However the rewards will be great when AI liberates us from repetitive activities and grants us more time for higher order things.

What Will This New Order Do to Your Data Center?

It’s far too early in the process to provide a definitive answer. What we do know is algorithms that recognize images and interactive voice applications are complex, energy rich processes.

Hence the luxury of being able to treat patients through digital assistance, and perform advanced what-if scenario modeling will come at a cost. That cost will include more complex and more expensive data processing centers.  

We therefore need to think ahead of the game and determine the right way to develop our infrastructure, in order to harvest the tremendous commercial opportunities AI offers for those sufficiently fleet-footed to keep up with it.

Is Cloud the Best Place to Do Our AI Work?

Cloud has become a popular host for artificial intelligence development. However, AI has developed into such a data-rich opportunity that it’s time to reconsider whether this is still the most cost-effective solution.

Michael Dell from PC Magazine believes, “when you modernize and automate your on-premises and co-location systems, such workloads will cost less in your own data centers than in the public cloud.”

He goes on to say that, “companies with a cloud-first strategy will ‘wake up’ and discover that it was more expensive”. The reason behind this is it may be cheaper to make a once-off purchase as opposed to renting heavy processing power at high hourly rates.

Which Hardware Would I Need if I Did the Work In-House?

Artificial intelligence developers deploy a new generation of neural network accelerators and processors. These evolved in tandem to meet a need for faster processing in hand-held devices. However since then specialized hardware developed to facilitate artificial intelligence and deep learning applications.

This functional separation meant basic central processing units (CPU’s) continued down the road of increasing capacity to perform complex computations. We move on to consider the most likely add-ons you may require, should you decide to do your AI development on your own hardware. These three technologies are briefly:

  • GPU’s for Data Sciences Matrix-Based Calculations
  • FGPA’s for Programmable Artificial Intelligence Hardware
  • ASIC’s for When Our System is Fully Stabilized

GPU’s for Data Sciences Matrix-Based Calculations

Graphical processing units (GPU’s) are dedicated hardware equipment for manipulating images and calculating their properties. They originally evolved for gaming applications. Later, AI scientists recognized their ability to accelerate other geometric calculations.

It was not long before they discovered they could apply the power of transforming polygons or rotating verticals in other situations. Nowadays the technology works equally well for the data science mathematics we benefit from today.

CPU’s excel at complex, repetitive math calculations, whereas GPU’s have a leading edge over data science’s simple, matrix-based ones. It follows you will need the power of these graphical processing units to open the door to artificial intelligence.

FGPA’s for Programmable Artificial Intelligence Hardware

Fluidity in the AI sphere means we are still evolving our deep learning frameworks. Hence it makes sense to use reconfigurable devices including field programmable gate arrays (FGPA’s).

FGPA’s facilitate inference, being reasoning moving from assumptions to logical consequences that follow when an argument holds. FGPA chips thus make it easier to evolve hardware, frameworks and software in parallel.

However, having this facility adds a further burden of power and performance costs. Therefore artificial intelligence developers may replace FGPA’s with cheaper, faster and more power-efficient alternatives after their applications bed down.

ASIC’s for When We Are Preparing to Go to Market Live

Application-specific integrated circuits (ASIC’s) can be ten-times more efficient than GPUs and FGPAs in the post development phase. That’s because these microchips are purpose made for particular applications.

Therefore, they can use solid state circuits, as opposed to connecting configurable blocks in FGPA scenarios. They therefore come into their own when we reach a point where it’s time to bed down our artificial intelligence system.

ASIC accelerators use optimized memory and less precise arithmetic to speed up calculations, and hence reduce power consumption. These lower costs make it easier to reach the net-earning stage and start clawing back expenses faster.

Thus Renting Versus Owning Can be a Staged Decision

Apple, Amazon, Google and other AI service providers have all produced their own customized hardware for completing deep learning and managing tasks. Their clients have to work with them to benefit from these gains.

It can make sense to use their GPU’s and FGPA’s while we are developing a particular AI application for our own use. That’s because we may never require the hardware again for a while.

However, the cost-logic may swing the other way as we move across to ASIC accelerators, because these solid-state purpose-made chips requite minimal human intervention.

Moreover, this two-stage approach switches the burden of temporarily-increased cooling across to the service provider during the developmental phase.  It’s therefore essential to include the power implications of rental-to-ownership migration in project costs.