Computer Optimization: Your Computer Is Faster Than You Think

“Welcome to the Jungle” (CPU design and Moore’s law)

“There’s plenty of room at the Top: What will drive computer performance after Moore’s law?”⁠, Leiserson et al 2020 (gains from End-To-End Princple-based systems design; AI scaling can continue even if semiconductors do not, particularly with self-optimizing stacks⁠; John Carmack notes, apropos of Seymour Cray⁠, that “Hyperscale data centers and even national supercomputers are loosely coupled things today, but if challenges demanded it, there is a world with a zetta[flops] scale, tightly integrated, low latency matrix dissipating a gigawatt in a swimming pool of circulating fluorinert⁠.”)

“Scalability! But at what COST?”⁠, McSherry et al 2015 (when laptops > clusters: the importance of good baselines)

“What it takes to run Stack Overflow”⁠, Nick Craver

“Anatomy of a hack: How crackers ransack passwords like qeadzcwrsfxv1331; For Ars, three crackers have at 16,000+ hashed passcodes—with 90% success”

“The Science of Brute Force”

Algorithmic progress:

• “Algorithmic Progress in Six Domains”⁠, Grace 2013
• “A Time Leap Challenge for SAT Solving”⁠, Fichte et al 2020 (hardware overhang: ~2.5× performance increase in SAT solving since 2000, about equally due to software & hardware gains, although slightly more software—new software on old hardware beats old on new. See previously .)
• 2019 Unlocking of the LCS35 Challenge
• “Measuring hardware overhang”⁠, hippke (“with today’s algorithms, computers would have beat the world chess champion already in 1994 on a contemporary desk computer”)

Memory Locality:

• “What Every Programmer Should Know About Memory”⁠, Drepper 2007
• “You’re Doing It Wrong: Think you’ve mastered the art of server performance? Think again.”⁠, Poul-Henning Kamp 2010 (on using cache-oblivious B-heaps to optimize Varnish performance 10×)
• “The LMAX Architecture”⁠, Fowler 2011

Deep Reinforcement Learning:

• Training A3C to solve Atari Pong in <4 minutes on the ARCHER supercomputer through brute parallelism
• “megastep helps you build 1-million FPS reinforcement learning environments on a single GPU”⁠, Jones
• “The Mathematics of 2048: Optimal Play with Markov Decision Processes”
• Bruteforcing Breakout (optimal play by depth-first search of an approximate MDP in an optimized C++ simulator for 6 CPU-years; crazy gameplay—like chess endgame tables⁠, a glimpse of superintelligence)
• “Sample Factory: Egocentric 3D Control from Pixels at 100,000 FPS with Asynchronous Reinforcement Learning”⁠, Petrenko et al 2020
• “Large Batch Simulation for Deep Reinforcement Learning”⁠, Shacklett et al 2021
• “Scaling Scaling Laws with Board Games”⁠, Jones 2021
• “Mastering Real-Time Strategy Games with Deep Reinforcement Learning: Mere Mortal Edition”⁠, Winter
• “Scaling down Deep Learning”⁠, Greydanus 2020
• “Podracer architectures for scalable Reinforcement Learning”⁠, Hessel et al 2021 (highly-efficient TPU pod use: eg solving Pong in <1min at 43 million FPS on a TPU-2048)

Latency/UI:

• “It’s the Latency, Stupid.”
• “Computer latency: 1977–2017”
• “Keyboard Latency”
• “Terminal Latency”
• “Local-first software: You own your data, in spite of the cloud”⁠/ Kleppmann et al 2019
• “Web Bloat”