Memory Panel Position Statements
Debate Title: Is Processing in/near memory worth the effort?
Moderator: Alaa Alameldeen (Intel)
Panelists: Onur Mutlu (ETH Zurich), Mattan Erez (University of Texas,
Austin)
- Alaa Alameldeen - Processing in memory has been a
popular topic for computer architecture research, on and off, for half a
century. Harold Stone’s 1970 paper, “A Logic-in-Memory Computer,” first
introduced the concept. Research picked up steam in the 1990s with
prototypes like Terasys from the Supercomputing Research Center, and
several academic research proposals. Many research efforts have investigated
moving various amounts of computations to or near memory. Research
in this area has increased recently, motivated by new memory designs that
support logic, and emerging memory technologies. Moving computations to
(or near) memory is expected to reduce energy, alleviate the memory
bandwidth bottleneck, and improve performance. However, processing in/near
memory would necessitate non-trivial hardware and software changes. Our
esteemed panelists will debate whether processing in memory is practical,
and whether it is worth the effort required to make it happen. As time
permits, the panelists will debate other memory-related research topics.
- Onur Mutlu (in support) - I argue that we have to
enable and exploit processing capability outside the
processors/accelerators and inside different levels of the memory
hierarchy, to minimize the huge data movement bottleneck that is plaguing
our computing systems. In fact, we have little choice but to do so, if we
want to make computing fundamentally more sustainable and energy-efficient.
Yes, it is not easy. Yes, there are a bunch of issues across the computing
stack that we need to re-examine, re-think and solve. Yes, your favorite
processor-centric mechanism will need to perhaps go away, or at least
re-thought within the new context... But, why not do it if it really makes
sense and enables a better future? No new paradigm or technology is easy,
and "processing data where it makes sense" is no different, with
likely huge payoffs. And, we can get there step by step, if only we can
overcome the biggest limiter of all: rigid mindsets.
- Mattan Erez (in opposition) - While I do research
on processing in/near memory myself, I argue that it is only an
interesting academic pursuit rather than a practical one. Yes, the
theoretical bandwidth and energy advantages of not moving data off of the
memory device are enticing. However, processing in memory requires
intrusive changes throughout the system while both the set of advantageous
operations and the set of applications that can benefit is limited. At the
same time, enabling memories to compute does not come for free, with
increased design times, complex testing that increases cost, and reduced
storage density. I argue that the costs outweigh the large, but
limited-scope benefits.
Back to CARD webpage