Why Computer Architecture Must Keep Improving

TL;DR

Better computer architectures must improve security, reliability, safety, energy efficiency, performance, latency, and predictability while supporting data-intensive fields such as machine learning, genomics, medicine, and health. Computer architecture covers systems with computation and memory, from visible devices to data centers, and its study reveals the principles and ideas behind their growing complexity.

Transcript

no you can hear me a little bit I don't know if this is working maybe we need to let's see is this better I think it's a bit better I don't want the echo though how about now good okay can people in the in on Zoom hear me if there are people yes okay somebody said they can hear us that's good always good to know okay it looks like this is live as w... Read More

Key Insights

• Computer architecture is the design of systems containing computation and memory capabilities, including devices people use directly and less visible infrastructure such as data centers. Studying it requires understanding the principles and ideas that make modern computing systems increasingly complicated.
• Fundamentally better computers must improve several distinct qualities, including security, reliability, safety, energy efficiency, performance, latency, and predictability. These qualities define different aspects of whether computing systems can be trusted, sustained, and used effectively.
• Trustworthy computing requires attention at both software and hardware levels. Software systems can produce unreliable answers, while hardware can also exhibit behavior that users cannot trust, so dependable computer design must address problems across the complete computing stack.
• Energy efficiency is essential to sustainable computing because current systems are described as highly energy hungry. Artificial intelligence systems are presented as especially inefficient, with their continued trajectory threatening sustainability goals and motivating research into data-centric and memory-centric architectures.
• Low latency means producing an answer or completing an operation quickly. It is particularly valuable when receiving a result sooner enables urgent action, as illustrated by a consequential DNA test whose result may guide needed treatment.
• Predictable latency means completing an operation at an expected time, which is not the same as minimizing delay. A tram leaving earlier than scheduled may technically have lower latency, yet it creates a worse experience for someone relying on its published departure time.
• Efficient data analysis is increasingly important in machine learning, genomics, medicine, health, and other domains. Computing systems must process and help interpret data efficiently so that people or applications can use the results to make decisions.
• Computer architecture education can progress from an introductory digital design and architecture course to a smaller research seminar and then a more advanced course. The seminar emphasizes reading cutting-edge research, presenting it, and developing presentation skills, while the advanced course builds on introductory material.

Questions & Answers

Q: What is computer architecture and what systems does it cover?

Computer architecture is the design of systems that have computation and memory capabilities. Its scope includes familiar computing devices as well as infrastructure that people may not directly see, such as data centers. The field examines how these systems work and how numerous principles and design ideas combine to create the complicated computing platforms used today.

Q: Why do computer architectures need continued research?

Computer architectures still fall short despite the many principles and ideas already incorporated into modern systems. Continued research is needed to generate new approaches for improving security, reliability, safety, trustworthiness, energy efficiency, sustainability, performance, latency, and predictability. The stated research mission is therefore not merely to build more computers, but to build fundamentally better computers across these dimensions.

Q: What qualities define a fundamentally better computer?

A fundamentally better computer is more secure, reliable, safe, trustworthy, energy efficient, sustainable, fast, and predictable. These qualities address different needs rather than representing a single measure of improvement. A system may return results quickly but unpredictably, or perform well while consuming excessive energy, so architecture must balance several objectives when defining and designing better computing systems.

Q: Why is energy efficiency important for artificial intelligence systems?

Energy efficiency matters because current computing systems are described as highly energy hungry, and artificial intelligence systems are characterized as extremely inefficient. Continuing along the same trajectory could make sustainability problems considerably worse. The architectural challenge is to support artificial intelligence without undermining climate and sustainability goals, including through data-centric and memory-centric approaches that reconsider how systems handle computation and data.

Q: What is the difference between low latency and predictable latency?

Low latency means receiving a result or completing an action quickly, while predictable latency means knowing when it will happen. The two properties are distinct and useful in different situations. A medical result may need the lowest possible delay, but transportation should follow an expected schedule because an unexpectedly early departure can increase a passenger's actual waiting time.

Q: When is predictable latency more useful than low latency?

Predictable latency is more useful when people or systems coordinate their actions around an expected completion time. The lecture uses a tram as an example: a departure earlier than scheduled may appear to reduce latency, but a passenger arriving according to the timetable can miss it and wait for the next tram. In that setting, predictable timing provides greater practical value than simply leaving sooner.

Q: When is low latency especially important?

Low latency is especially important when a quick result enables timely action. The lecture illustrates this with a DNA test that is important to a person's life. Receiving the answer quickly can allow the person to act sooner and obtain needed treatment. In such a case, reducing delay is valuable because the result directly affects how rapidly a consequential decision can be made.

Q: How can students continue studying computer architecture after the introductory course?

Students can continue through a computer architecture seminar and a more advanced architecture course. The seminar is described as a smaller course with about 20 students, where participants read cutting-edge research, present material, and develop presentation skills. The advanced course revisits and extends architectural topics while building on the principles introduced in the initial digital design and computer architecture course.

Summary & Key Takeaways

• Computer architecture concerns the design of systems that combine computation and memory, including everyday devices and data centers. The course begins at a high level to motivate the subject before examining the many principles and ideas involved in modern computing systems across roughly a full semester of lectures.

• The research mission presented is to build fundamentally better computers. Better systems should be secure, reliable, safe, trustworthy, energy efficient, sustainable, fast, and predictable. Current designs still fall short across these dimensions, creating a continuing need for research and new architectural ideas rather than simple acceptance of existing approaches.

• Low latency and predictable latency serve different needs. A critical DNA test benefits from returning results quickly, while public transportation is useful when it follows an expected schedule. Architecture must also support efficient data analysis in machine learning, genomics, medicine, health, and other domains where results inform consequential decisions.