Elevate your enterprise information technology and strategy in Transform 2021.
Of the concerns surrounding artificial intelligence these days — and no, I do not mean evil robot overlords, but more mundane things such as job replacement and security — perhaps none is more overlooked than cost.
This is understandable, considering AI has the potential to lower the cost of doing business in so many ways. But AI isn’t only costly to obtain and deploy, but it also requires a substantial quantity of energy compute electricity , storage, and energy to produce worthwhile returns.
Back in 2019, AI pioneer Elliot Turner estimated that coaching the XLNet natural language system could cost upwards of $245,000 — approximately 512 TPUs running at full capacity for 60 straight hours. And there is no guarantee it’ll produce usable results. Even a simple task like coaching an intelligent machine to solve a Rubik’s Cube could draw around 2.8GW of energy, regarding the hourly output of three nuclear power plants. All these are serious — although still debatable — amounts, considering that some estimates claim technology processes will draw over half of our global energy output by 2030.
Perhaps nobody understands this better than IBM, which has been at the forefront of the AI development — with varying levels of success –thanks to platforms like Watson and Project Debater. The company’s Albany, New York-based research lab has an AI Hardware Center that may be on the verge of unveiling some interesting effects in the drive to reduce the computational needs of training AI and directing its decision-making procedures, based on Tirias Research analyst Kevin Krewell.
A crucial development is a quad core evaluation chip recently unveiled in the International Solid-State Circuits Conference (ISSCC). The processor features a hybrid 8-bit floating-point format for coaching functions and 2- and 4-bit integer formats such as inference, Krewell wrote in a Forbes piece. This would be a substantial advancement over the 32-bit floating-point solutions that electricity present AI alternatives, but only if the right applications can be developed to produce the exact same or better results under these decrease memory and logic footprints. So far, IBM has been silent on how it intends to do this, even though the company has announced its DEEPTOOLS compiler, which supports AI model development and coaching, is harmonious with the 7nm die.
Qualcomm is also interested in driving greater efficacy in AI versions, with a specific focus on Neural Architecture Search (NAS), the means by which intelligent machines map the most efficient system topologies to achieve a given task. But since Qualcomm’s chips generally have a minimal power footprint to start out with, its focus is on developing new, more efficient versions that work smoothly within existing architectures, even at scale.
All for one
To that end, the company says it has adopted a holistic approach to modeling that stresses the need to psychologist numerous axes — like quantization, compression, and compilation — in a coordinated manner. Since all of these techniques complement each other, researchers have to deal with efficiency challenge out of their unique angle although not so that a change in 1 area disrupts gains in another.
When applied to NAS, the main challenges are decreasing high compute costs, improving scalability, and providing more accurate hardware performance metrics. Called DONNA (Distilling Optimal Neural Network Architectures), the solution provides an extremely scalable ways to define network architectures around accuracy, latency, and other requirements and then deploy them in real-world environments. The business is currently reporting a 20% rate increase over MobileNetV2 in locating highly precise architectures on a Samsung S21 smartphone.
Facebook also has a strong interest in fostering greater efficacy in AI. The Business re