9/13/2023 10:00:00 AM
Ghose exploring new possibilities for easy-to-program, practical PIM hardware thanks to Samsung, Sandia grants
Illinois CS professor Saugata Ghose believes new research opportunities will take his research interest in data-centric computer architectures and systems to the next level through industry-backed products.
Nearly 10 years ago, Illinois Computer Science professor Saugata Ghose began honing a research interest during his postdoctoral experience at Carnegie Mellon University that featured adding intelligence to computer memories and storage, particularly delving into the concept of processing-in-memory (PIM).
Due to a lack of intelligent capabilities in today’s memories, computers currently require costly operations for data computation, moving the data back and forth between memory and far-off processing hardware (CPUs, GPUs, and accelerators).
PIM introduces new hardware close to memory that can let the memory directly operate on the data, eliminating unnecessary data movement and energy waste for simple computation. In concept, PIM has tremendous potential because it tackles critical hardware inefficiencies exposed by the incredible increase in data usage during the social media upswing of the past several years.
However, the concept also undoes some key tenets that computer architectures have built off of for decades, which can disrupt the way programs are written and run and can make it difficult to design practical computers around PIM-capable memories. While other researchers focused on designing the hardware, Ghose stuck by his convictions and started to explore how one could easily program and integrate PIM.
Now, he is on the precipice of new possibilities for easy-to-program, practical PIM hardware. He has recently secured significant funding to continue his research. He has earned a five-year, $1 million grant from Samsung, with funds supported by the Samsung Memory Solutions Lab. Additionally, he has won a three-year, $273,000 grant with Sandia National Laboratories.
The way that computer hardware handles data has at a high level remained unchanged for the last 60-70 years. “Effectively, we keep data in one place and move it somewhere else to actually do any of the computation,” Ghose said. “This model worked really well when you didn’t have a ton of data. You could just move a little bit of data over, and you could do a whole bunch of computation. That was great, and it’s enabled a whole series of innovations over the last several decades that’s gotten us to where we currently are with computing.”
“But the problem now is that, increasingly, computational energy matters. And that’s really a lot of what’s been driving what I want to do. To me, there’s a need to take a step back and rethink a lot of the assumptions we have from the bottom up.”
To begin pairing the years of development Ghose has made on PIM with his students in the ARCANA Research Group, the Samsung grant offers what the professor called a “playground” for his programmability work.
Primarily, he said, the grant will connect his efforts with Samsung’s Smart SSD, which, according to the manufacturer’s website, is a “computational storage drive (that) leverages self-processing technology to reduce CPU utilization and data transfer bottlenecks.”
The opportunity is an exciting one in Ghose’s mind because Samsung has presented itself as a cutting-edge leader in developing technology advancements in data storage.
“Samsung has products that represent the leading edge of this type of processing in-memory and in-storage work,” Ghose said. “And they are actually looking to work with academics, involving them in their ecosystem in ways that I think is a bit unique. To me, this grant represents an awesome chance to be able to bring some of our work to life because a lot of the work that we do on our own ends up being in simulations.”
“We’re designing system models that are maybe 5-10 years out still. But this is a chance to be able to do things with products that are on the market today or will come to market over the next few years.”
Moving forward with Samsung, Ghose believes the ultimate goal is to spend these next five years working on a number of initiatives, including:
-Co-designing hardware and software that can allow programmers to make use of PIM efficiently.
-Creating software run times that can control the PIM data paths and redirect tasks to where the data
-Delivering compilers for processing-in-storage that allow programmers to continue writing in the
languages and the APIs they’re used to.
Simultaneously, his one-year grant with Sandia National Laboratories will delve further into the impact new processing in-memory architectures can have on machine learning and scientific computing.
Ghose said this goal with Sandia will be to build upon an emerging hardware technique being prototyped today. This technique, known as processing-using-memory (PUM), harnesses special electrical interactions to perform computation directly inside the actual data storage or memory chips themselves. He hopes to integrate existing analog PUM prototypes that have narrow use cases with a new digital version of PUM that his group has been developing with collaborators at Carnegie Mellon.
“PUM represents a new way to pack a whole bunch of data and compute into a really compact space,” Ghose said. “There’s been a lot of interest in this area focused on creating a sort of next-generation machine learning accelerator, and we’re certainly not the first to work in this space. But we’ve built upon the work in PIM and PUM that people have begun over the last 30 years, including the likes of our very own professor Josep Torrellas, with these digital versions of accelerators.”
He’s also found the partnership between his lab and Sandia to be a natural one.
First, this is due to the tie between this national lab and the University of Illinois. And second, it stems from his consideration that Sandia is a leader in this technology well before their association started.
Through this work and his new collaboration with Samsung, the long and sometimes lonely road Ghose has taken with students in his lab is overturning a new direction.
“I’m reminded of the feeling I had as a postdoc because it was clear that the work I was embarking upon was a bit unique and different. To come so far and understand now that we have companies like Samsung and a national lab like Sandia interested in funding this work solidifies why we took this approach to begin with, “ Ghose said. “I think this gives us a lot of confidence and a lot of energy. We have a chance to push some of our ideas forward and see the light of day.”