Discover the Future of Processors at Hot Chips 2023 Conference

Discover the Future of Processors at Hot Chips 2023 Conference

Table of Contents

1. Introduction to Hot Chips
2. Day 1: Tutorial Day
   • Machine Learning Inference
   • Chiplets and Universal Chiplet Interconnect Protocol
3. Day 2: Conference Day 1
   • Exciting Directions for Machine Learning Models
   • Processing in Memory
   • CPU Session
4. Day 3: Conference Day 2
   • Poster Lightning Talks
   • Good Hardware for Deep Learning
   • Machine Learning Training
   • Interconnects
   • FPGAs and Cooling
5. Conclusion

Introduction to Hot Chips

Hot Chips is an annual conference that brings together a wide range of individuals interested in computing, microarchitecture, and the latest chips in the market. From beginners in computing to leading industry engineers, Hot Chips offers a platform for the community to discuss and showcase the advancements in chip design. This year, the conference will be held from August 27th to 29th at Stanford University, adopting a hybrid approach with both in-person and online participation.

Day 1: Tutorial Day

The conference kicks off with Tutorial Day, providing attendees with an in-depth understanding of two key topics: machine learning inference and chiplets. In the machine learning inference session, industry experts will explore the importance of optimizing inference models for performance and cost efficiency. Talks by Micah Vilmo from Nvidia, Microsoft on deep speed, and Qualcomm on inference at the edge will delve into the advancements and challenges in this area.

The chiplets session centers around Universal Chiplet Interconnect Protocol (UCI), a standard that is gaining traction in the industry. Intel and AMD will provide an overview and usage models of UCI. Talks by Google, Intel, and Nvidia will address the electrical, form factor, and software aspects of chiplets. These Sessions allow attendees to gain insights into this emerging trend and its potential impact on chip design.

Day 2: Conference Day 1

The first day of the conference focuses on various topics, including machine learning models and implications for computing hardware, processing in memory, and CPU advancements. Jeff Dean and Amin Vahdat from Google will deliver a keynote on exciting directions for machine learning models, exploring the implications for computing hardware.

The processing in memory session will delve into the integration of compute and memory, optimizing data movement and processing. SK Hynix will discuss memory-centric computing, followed by presentations from Samsung on AI cluster systems and Kim on CX1-based processing near memory for Transformer-based LLMs. These talks aim to shed light on the advancements in memory-centric computing and its applications in machine learning.

The CPU session will offer insights into the latest developments in CPU architecture. Talks by Magnus Bruce from Arm on Omniverse V2, Kaitrostor from AMD on the next-generation Zen4, and Ventana Micro showcasing their Veyron V1 CPU will provide attendees with in-depth knowledge about these powerful processors and their applications.

Day 3: Conference Day 2

Day 3 of the conference features lightning talks, covering a wide range of topics such as Azure and AWS, 130 nanometer open source ISE, multi-chip deep learning accelerators, and more. These presentations offer a quick glimpse into ongoing research and industry developments.

The conference continues with talks on good hardware for deep learning by Bill Dally from Nvidia and the machine learning training session, which will explore Google's machine learning supercomputer, Cerebras' Wave Scale cluster, and Nvidia's North Pole neural inference machine. These talks aim to provide insights into cutting-edge hardware solutions for machine learning training.

The interconnects session will focus on networking advancements, including Nvidia's resource-fungible networking processing ASIC, Lightelligence's hummingbird low latency computing engine, and Intel's first direct mesh to mesh photonic Fabric. These talks highlight the latest developments in optimizing networking performance for different applications.

The conference concludes with talks on AMD's next-generation FPGA design, Intel's agile X9, and high-performance cold plates for data center thermal management by Fabric8 Labs. These presentations explore the advancements in FPGA architectures, cooling technologies, and their implications for data centers.

Conclusion

Hot Chips offers a unique platform for industry professionals and enthusiasts to come together and explore the latest advancements in chip design. From machine learning inference to chiplets, CPUs, and interconnects, the conference covers a wide range of topics that are shaping the future of computing. By attending Hot Chips, participants can gain valuable insights from industry leaders, discover emerging trends, and network with like-minded individuals passionate about computing and microarchitecture.