IBM AIU—A System On A Chip Designed For AI
IBM AIU—IBM’s First System On a Chip Designed For AI
IBM recently announced its first system-on-chip designed specifically to run deep learning AI models that it’s calling Artificial Intelligence Unit (AIU).
IMB AIU is an application specific integrated circuit (ASIC) with AI optimized design features.
Although it’s IBM’s first AI chip, it’s joining the ranks of other AI focused ASICs like Google’s TPU, and AWS’s Trainium chips.
CPU, GPU, FPGA, TPU, and now IBM AIU—What’s an AI Model to Use?
Central processing units (CPU) are the foundation of modern computers and although very powerful for many tasks.
However, CPUs are painfully slow at running modern AI deep learning models.
Deep learning models run best on processors that allow for massive numbers of threads running in parallel.
That’s why graphics processing units (GPU) are commonly used for running AI models.
While a high-end CPU might have dozens of cores, a high-end GPU can have thousands of cores, providing for that massive parallelization needed to efficiently run AI workloads.
Field-programmable gate arrays (FPGA) are also widely used for AI processing.
FPGAs are integrated circuits (IC) that can be reprogrammed to perform specific functions.
FPGAs that are programmed to perform specific AI tasks are typically not as fast as ASICs.
FPGA’s, however, have the flexibility of being re-programmed to optimize for a particular AI algorithm.
Less Precision Delivers Speed and Accuracy
IBM uses what it calls “approximate computing” in its chip design, using a range of smaller bit sizes for data.
This approach sacrifices precision, although IBM is claiming for AI this is not a concern.
The smaller bit formats significantly reduce the time needed to move data in memory as well as the processing time needed to train and run AI models, all without sacrificing accuracy.
IBM Is Focusing On AI Hardware
This chip was the manifestation of five years of design work at IBM Research AI Hardware Center in Albany, NY.
It contains 32 processing cores containing 23 billion 5 nm transistors and was based in the AI accelerators built into IBM’s Telum chip.
The IBM Research AI Hardware Center has an aggressive roadmap to improve AI hardware efficiency by 2.5x each year.
They aim to train and run AI models 1,000x faster in 2029 than they could in 2019.
Bottom Line
IBM is joining the ranks of Google and Amazon Web Services in creating its own AI-optimized ASIC.
This is good news for enterprises, as more and more technology providers are incorporating AI into their software applications, and as hardware capabilities continue to multiply.
AI algorithms and model capabilities will continue to improve as well.
Upcoming Webinars
Conversational AI—Putting Digital Labor to Work
Conversational AI continues to evolve and improve throughout 2022 and is being adopted broadly across the enterprise.
Significant improvements in AI algorithms and the hardware they run on have combined to enable technology providers to deliver solutions that in many cases are changing how work is done.
Join Aragon Research’s Sr. Director of Research, Craig Kennedy, on November 9, 2022.
He will discuss the current state of Conversational AI and why it is now a must have technology for the digital enterprise.
This webinar will cover:
- How did chatbots get so darn smart?
- What are some of the key trends in conversational AI in 2022?
- Why should enterprises employ conversational AI?
This blog is a part of the Digital Operations blog series by Aragon Research’s Sr. Director of Research, Craig Kennedy.
Missed an installment? Catch up here!
Blog 1: Introducing the Digital Operations Blog Series
Blog 2: Digital Operations: Keeping Your Infrastructure Secure
Blog 3: Digital Operations: Cloud Computing
Blog 4: Cybersecurity Attacks Have Been Silently Escalating
Blog 5: Automation—The Key to Success in Today’s Digital World
Blog 6: Infrastructure—Making the Right Choices in a Digital World
Blog 7: Open-Source Software—Is Your Supply Chain at Risk?
Have a Comment on this?