23. Februar 2021, 10:00 Uhr | Steve Pawlowski and Prasad Alluri
AI applications place high demands on memory.
AI applications are driving demands for additional data storage. Data storage units must hold increasing amounts of data and output large blocks of these data to the processor very quickly. As a result, the structures of memories in computer systems for AI will have to change.
AI will become more accurate and more ubiquitous, and, we'll start to see it filling in more gaps for simple tasks where people would traditionally say »Why would I ever use an AI algorithm for that?« For example, grocery stores might tap AI-enabled cameras that periodically check to see if a shelf is empty and if so, alert a clerk to restock.
In a post-COVID world, we’ll see more businesses adopting AI for use cases like these to create these contactless experiences. We’ll also see AI moving into infrastructure such as data centers and 5G base stations as neural network algorithms become more adept at workload and system error correction and recovery.
You’ll have AI algorithms for simple, singular tasks and sophisticated algorithms for mapping the human brain. AI this agile requires a significant amount of memory bandwidth, and when you add memory bandwidth, you’re adding power – researchers have found that training a single AI model can emit 5x the lifetime carbon emissions of an American car (including manufacture) .
These immense power requirements aren’t sustainable if AI is to be pervasive in our society both in terms of the data center and our planet. In the next several years, we’ll see ecosystem players exploring new ways of powering AI in an energy-efficient manner. That could be taking AI architecture and moving it closer to the memory via stacking memory in a 3D package on top of the logic. Beyond the hardware, today there's a lot of bandwidth – and data – we're leaving on the table, meaning huge amounts of energy wasted.
So in the years ahead, we’ll see a rise in technologists analyzing how to finetune AI algorithms to get as high performance as possible, but with extremely stingy power levels. Basically, it’s squeezing insights out of every bit of data to ensure it’s not going to waste.
The increase in AI means that its increasingly important that edge computing is near 5G base stations. So soon, in every base station, every tower might have compute and storage nodes in it. And there are lots of startups that are focused on building edge data centers that look like transport containers that sit in metro areas to enable content – like your Hulu videos – to be closer to the consumption. We’ll see the adoption of these edge data centers in the next few years, as enterprises and consumers look to tap massive amounts of data for insight and faster services closer to the source.
In 2021, look for more usage of object stores, for storing structured and unstructured data, files, blocks, objects – all in one repository. AI’s large data sets need proximity to where processing is happening. So, rather than viewing it as a large cold store, object stores are going to be able to do AI-type workloads, which means large sets of data can be accessed with high bandwidth and low latency.
As a result of the rise of data-intensive AI workloads, we’ll see the need for high-performance NVMe storage also increase, since this high-performing object store resides on flash-based storage, as opposed to the traditional, cold storage. This could lead to faster adoption of Gen4 NVMe SSDs to enable more powerful object store for AI.
 Hao, K.: Training a single AI model can emit as much carbon as five cars in their lifetimes. MIT Technology Review, 6.6.2019, www.technologyreview.com/2019/06/06/239031/training-a-single-ai-model-can-emit-as-much-carbon-as-five-cars-in-their-lifetimes.
is corporate vice president of advanced computing solutions and emerging memory solutions at Micron. He is responsible for defining and developing innovative memory solutions for the enterprise and high-performance computing markets. Prior to joining Micron in July 2014, Mr. Pawlowski was a senior fellow and the chief technology officer for Intel’s Data Center and Connected Systems Group. Steve earned bachelor’s degrees in electrical engineering and computer systems engineering technology from the Oregon Institute of Technology and a master’s degree in computer science and engineering from the Oregon Graduate Institute. He also holds 58 patents.
is the vice president of corporate strategy at Micron. Prior to joining Micron, he held leadership roles at Intel in engineering and business, ranging from memory, wireless, and optical product lines. He started his career as a researcher in non-volatile memory and advanced transistor development at Motorola research labs. Prasad holds a Ph.D. from Arizona State University and an MBA from The Wharton Business School.