Beyond Moore's Law: Software, Algorithms, and Architecture as New Performance Drivers

When Moore's Law reaches its limits, the question arises whether humanity’s computing power will be locked. A recent Science paper co‑authored by researchers from MIT, NVIDIA, and Microsoft argues that it will not; future performance gains will come from the top of the computing stack—software, algorithms, and hardware architecture.

Software performance engineering can dramatically improve efficiency by refactoring code, eliminating software bloat, and tailoring programs to specific hardware features such as parallel processors and vector units.

An illustrative example multiplies two 4096×4096 matrices. A naïve Python implementation (Version 1) takes about 7 hours on a modern machine, achieving only 0.0006 % of peak performance. Re‑implementations in Java (Version 2) and C (Version 3) speed the task up 10.8‑fold and 4.4‑fold respectively, while further optimizations—parallel execution on 18 cores (Version 4), cache‑aware coding (Version 5), vectorization (Version 6), and use of Intel AVX instructions (Version 7)—reduce the runtime to 0.41 seconds, a 60 000‑fold improvement over the original Python code.

Algorithmic advances also play a crucial role. Since the late 1970s, improvements in maximum‑flow algorithms have matched the gains from hardware scaling, though progress is uneven and eventually diminishes. Future breakthroughs are expected from new problem domains (e.g., machine learning) and algorithms designed to exploit emerging hardware.

Hardware architecture discussion begins with Dennard scaling, which kept power density constant as transistors shrank, enabling performance to double roughly every two years. Dennard scaling ended around 2005, leading to the multicore era. Performance data (SPECint, SPECint‑rate) show that after 2004 parallelism drove a 30‑fold increase in parallel workloads, while single‑thread performance lagged.

Two main strategies for future hardware are simplification (replacing complex cores with simpler, more numerous ones) and domain specialization (custom hardware for specific applications, such as reduced‑precision floating‑point for machine learning).

In the post‑Moore era, gains from silicon process improvements will be modest, while top‑down improvements in software, algorithms, and streamlined hardware will dominate, though they will be sporadic and opportunity‑driven. Emerging technologies like 3D stacking, quantum computing, photonics, superconducting circuits, neuromorphic computing, and graphene chips hold long‑term potential but are not yet competitive with silicon.

For more details, see the original Science article: https://science.sciencemag.org/content/368/6495/eaam9744