In a recent press release from the ETH Zurich, Onur Mutlu, a widely recognized computer scientist at the ETH, previously Carnegie Mellon University, share his impressions and analysis of the UPMEM PIM architecture.
In part, we can read that “Mutlu and his research group have characterized, analyzed, and tested the new (UPMEM) system and compared it with a previous state-of-the-art system with CPUs. They have learned that the novel system makes computing up to 23 times faster and five times more energy efficient.”
PIM systems enable a fundamentally more efficient approach to computing, unlike current standard systems where the processor made up of CPUs takes centre stage and the memory units are located far away from it. In such systems, the movement of data between the processor and memory nowadays limits the speed and efficiency of data processing, even though the processor itself is very powerful. And moving data requires the most energy – much more energy than the actual computation. “Accessing main memory just once consumes one hundred to a thousand times more energy than a complex addition,” Mutlu says.
Mutlu and his colleagues have tested the novel system for applications in the fields of data analysis, databases, bioinformatics, image- and video analysis, and neural networks, among others. The PIM-system is best suited for workloads requiring little communication between DPUs (e.g. database and image applications)and primarily simple arithmetic operations (e.g. video analytics or data filtering). “We expect that as these systems evolve, they will become even faster and more energy efficient, and their applications will become even more diverse,” Mutlu reckons.”
Wide potential applications
What kind of companies could use this new computing system? “Energy efficiency and sustainability should be key goals for any industry”, Mutlu believes. “But companies that already have an idea of how they will use the new PIM hardware will benefit immediately”, he says. “If their workloads are a good fit for the new architecture, performance and energy efficiency will improve greatly.”
Data centers could use the PIM system already.
“ Every software company should also consider it in order to be prepared for the future,” Mutlu says.
Also, PIM is a substrate that makes embedded systems such as AR/VR glasses, drones, self-driving cars and others much more efficient. “Companies operating in these areas should think about how they could benefit from PIM in their systems.”
Such conclusions follow extensive research works from Onur’s team, known as the SAFARI research group. For the past 2 years, the team has proposed a 360 degrees overview of the UPMEM PIM architecture and its benefits across a full range of algorithms, including comparisons to CPUs and GPUs. This extensive paper led by SAFARI lead scientist Juan Gomez-Luna can be found here.
The SAFARI team is now pushing further several key PIM implementations for:
Machine Learning core algorithms
DNA alignment algorithms
Use of Sparse Matrix multiplication
Relational algebra
Graph pattern algorithms
A framework for Mapreduce
Look up tables applications