HPC Asia 2025 invited talk: dataflow under the von Neumann machine
Xian-He Sun gave the third invited talk at HPC Asia 2025 in Hsinchu, Taiwan, on February 20, 2025. The program lists it as "Dataflow under the Von Neumann machine: A Destructive New under Existing Systems". Sun argued that the memory wall, not raw compute, now limits AI systems, and presented a data-centric approach to it.
The memory wall returns
The success of deep learning depends on processing vast amounts of data, and computing systems are struggling to supply it. Sun argued that this demand has brought researchers back to the memory-wall problem: processors can compute far faster than memory and storage can deliver data to them. His answer is a concept he calls dataflow under the von Neumann machine, which keeps the conventional architecture but organizes computation around the cost of moving data.
Two models: C-AMAT and LPM
The talk built the concept on two pieces of GRC research.
- C-AMAT (Concurrent Average Memory Access Time) extends the classic average memory access time model to modern memory systems that serve many accesses at once. It counts how many hits and misses are in flight together and separates the misses that actually stall the processor from those hidden behind other work. This quantifies the benefit of concurrent data access and shows how data locality and concurrency relate.
- LPM (Layered Performance Matching) applies C-AMAT layer by layer. It tunes each level of the memory hierarchy so that the rate at which it can supply data matches the rate at which the level above requests it, so no layer becomes the one that stalls the rest.
With these two tools in place, Sun formally introduced dataflow under the von Neumann machine.
From software to hardware
Sun then turned to GRC's work on I/O systems, in particular the Hermes multi-tiered I/O buffering system. Hermes uses the LPM framework to decide how data moves between memory, flash and disk tiers, and Sun described it as a clear success. That raised the question at the center of the talk: Hermes is a software solution for I/O systems, so what is the practical hardware solution for memory systems? He closed with open problems and ideas for computer system designs for AI and big data that would ease the memory wall.
The venue
HPC Asia 2025 was hosted by Taiwan's National Center for High-performance Computing (NCHC) under the theme "Chip-based Exploration and Innovation for HPC". Sun's talk ran from 13:30 to 14:20 and was chaired by Taisuke Boku.
Learn more
- HPC Asia 2025 program (opens in a new tab)
- HPC Asia 2025 speakers, with the talk abstract (opens in a new tab)
- Concurrent Average Memory Access Time (IEEE Computer, 2014)
- LPM: A Systematic Methodology for Concurrent Data Access Pattern Optimization (IEEE TPDS, 2019)
- ACA 2020 keynote, where Sun first presented dataflowV
- Hermes project page
- Xian-He Sun's profile
