Computer Architecture Today

Informing the broad computing community about current activities, advances and future directions in computer architecture.

Welcome to the ASPLOS 2024 trip report. This year, ASPLOS 2024 was held in sunny San Diego from April 27- May 1, 2024 and was the largest ASPLOS ever. With ~800 attendees across 21 countries, ASPLOS 2024 was nearly double the size of ASPLOS 2023. General Chairs Nael Abu-Ghazaleh and Rajiv Gupta worked tirelessly over the past year to accommodate the growing ASPLOS 2024, including renegotiating the venue contracts multiple times! 

For the second year in a row, ASPLOS 2024 saw a 50% increase in number of submissions (~900), more than doubling the number of submissions for ASPLOS 2022. Program Chairs Dan Tsafrir and Madan Musuvathi were tasked with heroically managing this review process, involving ~240 PC members, 3,624 reviews, 12,655 comments, and ~50 pages in messages from the program chairs detailing the review process with lots of additional statistics. This year’s jam-packed program consisted of 10 tutorials and 12 workshops over the weekend, with 4 keynotes, a debate, WACI and 193 papers across 44 paper sessions over 3 full days.

Workshop and Tutorials

ASPLOS kicked off with 2 days of workshops and tutorials attended by over 350 attendees. A diverse set of interactive tutorials provided hands-on experiences with various tools, frameworks and platforms relating to PyTorch, open-source RISC-V processors, MLIR for Ryzen AI devices, OpenHarmony OS, Quantum applications, Quantum Circuit, Visual Computing, NPU Generators, Instrumentation frameworks for DNN, and detection of microarchitectural attacks.

The workshop program featured many well-established workshops and new workshops on emerging topics. New workshops this year introduced us to optimization of tensor methods (XTensor), open-source hardware design and EDA tools (OpenCircuit), computer systems for biomedicine (BioSys), and ethical computer systems (HotEthics)

Returning workshops covered topics such as computer platforms for autonomous vehicles, unary computing, composable disaggregated systems, and accelerator design. Notably, the Workshop on Cloud Intelligence / AIOps reaches half a decade and EMC2 approaches a decade with its 9th edition. Last but not least, the 6th Young Architect Workshop (Yarch) provided valuable mentoring to many early-stage computer architects. Yarch featured personal and captivating keynotes from Samira Khan and Adrian Sampson, and highlighted 19 lightning talks from student attendees.  

Keynotes and Debate

The first keynote was by Dr. Amin Vahdat, who is a Fellow and VP of Engineering at Google. Amin’s keynote focused front and center on the seismic shift that generative AI will have on datacenter systems, and how, in fact, systems innovation will play a pivotal role in shaping the AI future. He highlighted some of the innovations in the latest generations of Google’s TPUs, such as high-bandwidth interconnect for parameter distribution, liquid cooling, optical circuit switching, specialized data representations, and HBMs. Dr. Vahdat highlighted that these advances are still insufficient to meet the insatiable demand for ML computing, which is growing exponentially at the rate of 10x/year. In a bid to find the next 1000x, he encouraged the ASPLOS audience to think end-to-end and optimize for system goodput, power, and reliability through both algorithmic innovation and software/hardware codesign. He also highlighted the need to rethink traditional metrics, such as performance per TCO (total cost of ownership) as these metrics are built on the assumptions that there is enough datacenter capacity to house new computers and that power is infinitely available, both of which are no longer true. Power, in particular, appears to be a key bottleneck in scaling ML compute, and he encouraged the audience to think of innovations to get more out of each watt, including novel telemetry and control systems that do not yet exist. A recurring theme in his talk was Google’s commitment to sustainability and he highlighted the need to define better metrics that capture the impact of carbon emissions. 

The second keynote was by Emmett Witchel, who is a Professor of Computer Science at the University of Texas at Austin. Prof. Witchel opened his keynote with a lesson on Gustav Freytag’s pyramid and how many storylines in CS research follows this familiar trend. The keynote then focused on the story of CXL (which allows hosts to talk to memory over PCIe bus) and the challenges and opportunities it brings. The keynote concluded with advice on the purpose of research and how it is fundamentally driven by insight. He mentioned various notable research insights, including how these insights arise within groups, since research is innately a social activity.

The third keynote was by Dr. Tamar Eilam, who is a Fellow and Chief Scientist for Sustainable Computing in the IBM T. J. Watson Research Center. Dr. Eilam gave a fascinating talk on measuring and reducing carbon footprint at every layer of the computing stack and across the entire lifecycle of computing. She touched on interesting design tradeoffs that come to light when one starts thinking about embodied carbon, be it in semiconductor manufacturing, or in OS design. For example, she highlighted that specialized ASICs increase operational efficiency, but also increase embodied carbon footprint due to the carbon-intensive process of chip design and manufacturing. For those interested in learning more, she recently wrote a SIGARCH blog post on the subject, and her slides are available here.

The fourth keynote was by Nafea Bshara, who is a Vice President and Distinguished Engineer at Amazon Web Services (AWS). Nafea gave an insightful talk into “one of the largest chip companies you’ve never heard of”. The keynote details many of the unconventional full-stack optimization and design of the Tranium processor for ML workloads. For example, Tranium adopts a CISC design with large cores that offloads almost all hardware decisions to compilers, including cache management, register renaming, and hazard prevention. Many other design decisions were enabled by the fully vertically integrated nature of AWS where you can optimize not just the processor and compilers, but everything in between up to the datacenter-level, such as hypervisors, cooling, electricity, etc. 

Monday evening included a special debate on “Should everyone work on machine learning/AI?”, organized by Poulami Das and Joseph Devietti. Debate participants took turns making a case for, or against, whether one should work on ML/AI. Debating in favor of everyone working in ML were Amir Yazdanbaksh (Google DeepMind), Vijay Janapa Reddi (Harvard) and Charith Mendis (UIUC). Many arguments for working on ML-related problems revolve around the ubiquity of ML in products and the opportunity in new research problems that it opens. 

Debating against the need for everyone to work on ML were Tamar Eilam (IBM), Moin Qureshi (Georgia Tech) and Josep Torrellas (UIUC). These arguments revolve around the need for research diversification rooted in financial theory, the need to foster multidisciplinary research and following one’s research passion, and that ML/AI depends on many foundational areas such as energy-efficiency, security, and systems work. 


Always a highlight of ASPLOS, WACI this year featured 4 funny and provocative talks with a keynote from Prof. Zachary Tatlock (University of Washington). Guy Wilks (UCSB) introduces new security vulnerabilities discovered with the recent Alaska Airlines Flight 1282 event. BoltBleed discusses security vulnerabilities where devices with sensitive data may literally fall into the wrong hands. Nitesh Narayana, Gondlyala Sathya and Abhijit Das (UPC) gave a remote talk on how wireless network on-chip can operate at good speed to enable direct access between RF and DRAM. Michael Roitzsch (Barkhausen Institut) discussed how bureaucracy is entwined in systems designs, such as microkernels, and how we need to find the right amount of bureaucracy when designing computer systems. Finally, Deniz Altınbüken and Martin Maas, (Google DeepMind) discussed how generative AI can benefit the research process by generating research ideas and assisting with papers and talk generation, along with a fully generative demo presentation of Quantum Entanglement for Distributed Garbage Collection.  

Business Meeting

The business meeting was held Monday evening with various presentations from conference Chairs and a presentation of ACM E&P Committee with updates on ethics and plagiarism violations. The General Chairs provided updates on registration statistics and conference budget. The Program Chairs presentation led to a lively discussion on how to handle the challenges of growing submissions to ASPLOS and the challenges of reviewing over multiple cycles.  A particularly noteworthy announcement: As many of you may recall, ACM TOCS sunsetted at the end of 2023. ( Through hard work by the community, ACM TOCS has been revived. Please submit your best work to make ACM TOCS a continued success. 

Finally, ASPLOS 2025 will be co-located with Eurosys for the first time ever from March 30 – April 3, 2025, tentatively in Rotterdam, the Netherlands. Hopefully we’ll see you all there!

About the Authors: 

Akanksha Jain is a Software Engineer at Google. Her research interests include CPU microarchitecture, datacenter performance, HW/SW co-design, and ML for systems. 

Daniel Wong is an Associate Professor in the Electrical and Computer Engineering Department at the University of California, Riverside. His research interests include energy-efficient computing, data center architectures, and GPGPUs.

Disclaimer: These posts are written by individual contributors to share their thoughts on the Computer Architecture Today blog for the benefit of the community. Any views or opinions represented in this blog are personal, belong solely to the blog author and do not represent those of ACM SIGARCH or its parent organization, ACM.