BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Stash: Have Your Scratchpad and Cache it Too
DTSTART:20150427T140000
DTEND:20150427T153000
DTSTAMP:20260407T063936Z
UID:e61a9be301b6b78f7e38fef96f526fa6cf705ece03852f098983a93e
CATEGORIES:Conferences - Seminars
DESCRIPTION:Matt Sinclair is a fourth year PhD student at the University o
 f Illinois at Urbana-Champaign where he works in Professor Sarita Adve's r
 esearch group. His research focuses on building efficient memory hierarchi
 es for heterogeneous systems. He is a winner of a Qualcomm Innovation Fell
 owship. \nHeterogeneous systems provide a natural path to energy efficienc
 y. Given that the memory hierarchy is expected to be a dominant consumer o
 f energy in these systems\, efficient data movement is vital. However\, th
 e memory hierarchies of modern heterogeneous systems have multiple ineffic
 iencies such as loose coupling between compute units\; specialized\, priva
 te memories that only exist in local address spaces\; poor support for syn
 chronization primitives\; and complex memory consistency models. My work p
 roposes an efficient and easier-to-program heterogeneous memory system whe
 re specialized memory components are tightly coupled in a unified and cohe
 rent address space. Underlying the system is a low overhead\, software-dri
 ven hardware coherence protocol called DeNovo. This talk will focus on app
 lying these ideas to a heterogeneous system with CPUs and GPUs employing s
 cratchpads and caches. Specifically\, we propose a new memory organization
 \, stash\, that combines the benefits of caches and scratchpads without in
 curring the downsides of either. Like a scratchpad\, the stash does not in
 cur overheads of tags\, TLB accesses\, and conflict misses. Like a cache\,
  the stash is globally addressable and visible and allows for data reuse a
 cross compute kernels. With these benefits\, the stash shows higher perfor
 mance and lower energy than either a cache or a scratchpad\, while enablin
 g new applications for heterogeneous systems. Time permitting\, I will als
 o show how the use of the DeNovo protocol permits simpler memory consisten
 cy models than those motivated by current coherence protocols for heteroge
 neous systems.Cookies and drinks will be available before the talk as from
  13:45.
LOCATION:BC 420 https://plan.epfl.ch/?room==BC%20420
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR