Saved links with concise summaries.
Instruction-level parallelism allows a single CPU core to execute multiple instructions simultaneously by overlapping the stages of instruction processing, such as fetching, decoding, and execution. This technique, often implemented through pipelining and superscalar architectures, improves throughput without necessarily reducing latency. Modern CPUs achieve this by executing instructions out of order and utilizing multiple execution units to maximize resource utilization.
2024-03-24
The article explains the Leaky Bucket algorithm for rate limiting, contrasting it with the Token Bucket approach. It details the 'pure' bucket method and its queue-based variant, providing a Java implementation and discussing trade-offs like handling bursty traffic and resource utilization.
Linus Torvalds explores his fascination with programming, comparing it to physics and mathematics. He describes the process as creating one's own world within the computer, blending art and engineering. The article emphasizes the satisfaction of finding elegant, non-brute-force solutions to complex problems.
2026-04-23
The author optimized a Rust program deserializing AWS SDK models, saving 475 MB of memory. By changing struct layouts to use Option<Box<Struct>> instead of Option<Struct>, the author ensured that empty structs only take up a pointer size in memory. This technique leverages Rust's memory layout specifics and heap allocation to reduce fragmentation and footprint.
The author describes a method to improve Large Language Models by duplicating a block of seven middle layers without changing weights or training. This experiment reveals an 'LLM Neuroanatomy,' suggesting that internal reasoning occurs in contiguous, indivisible functional circuits rather than independent, interchangeable operations.
ASML is the sole manufacturer of Extreme Ultraviolet (EUV) lithography machines, the technology essential for producing the world's most advanced semiconductors. The company succeeded by betting on an unproven technology and collaborating with global partners to solve complex physics problems. Its machines now serve as the primary bottleneck in the global semiconductor supply chain.