Unified Memory: What Apple Silicon Changed for App Performance
Apple Silicon puts CPU, GPU, and Neural Engine on one memory fabric so apps skip a class of copies — and pick up new rules for residency, bandwidth, and Metal.
Category
91 articles · 314 min of reading
Apple Silicon puts CPU, GPU, and Neural Engine on one memory fabric so apps skip a class of copies — and pick up new rules for residency, bandwidth, and Metal.
Micikevicius et al. (2018) trained deep nets with half-precision arithmetic using loss scaling and fp32 master weights. The recipe still sitting under AMP, bf16, and most large-model trainers.
Huang et al. split a deep net across accelerators, pump micro-batches through the pipeline, and recompute activations to fit giant models. Bubble time, batch size, and when tensor parallelism is the better split.
Leviathan et al. use a small draft model to propose token blocks and the large model to accept or reject them in parallel. Same distribution, fewer serial large-model steps, and the acceptance-rate math that decides whether you win.
TensorRT lowers trained networks into fused, precision-calibrated kernels so production inference can use Tensor Cores instead of running training graphs unchanged.
NVIDIA's CUDA turned GPUs from graphics pipelines into a SIMT programming model that scientific computing and later neural networks could actually ship on.
Databricks built Photon as a native, vectorized engine so SQL and Spark workloads could leave interpreted row-at-a-time execution on the table.
RFC 8297: 103 with Link preloads while the origin still computes 200, CDN support, and why a slow TTFB still benefits from hinted CSS.
animation-timeline: scroll() and view(), ranges, and how to animate on scroll without a rAF loop that fights the compositor.
How to interpret width versus height, distinguish on-CPU from off-CPU, and use allocation flames to catch the copy you cannot see in the source.
How Redis kept a single-threaded command loop to make operations atomic and latency predictable — and where that design still bites under big keys and slow commands.
How the Speculation Rules API differs from rel=prefetch, when prerender is safe, and how to avoid double pageviews and stampedes on search-as-you-type.
Link rel=prefetch, Speculation Rules, and why prerendering URLs with tokens or personalization is a data leak as much as a performance win.
spring.threads.virtual.enabled, Tomcat's protocol, synchronized pinning on old JDKs, and when a platform-thread pool still belongs in the architecture.
What the compiler memoizes, which useMemo calls become redundant, and the purity rules that make it bail out of a component.
How React matches lists between renders, what keys actually identify, and the local-state bugs that index keys create.
What Zone.js was doing on every async hop, how zoneless + signals mark dirty components, and the remaining places you still need to tell Angular that data arrived.
How HotSpot decides an object never leaves a method, what scalar replacement actually removes, and how to read JFR when allocation still spikes.
JEP 454: Arena, MemorySegment, Linker, and why bounded lifetimes beat ByteBuffer folklore for off-heap and native calls.
Non-destructive hydration, event replay, and the mismatches that force a full client rebuild — plus incremental hydration for heavy islands.
React Hook Form, native constraint validation, and why defaultValues plus uncontrolled inputs beat a controlled 80-field form.
How B-trees update pages in place and LSM trees turn writes into sorted flushes, with the read/write/space trade-offs that show up in MySQL, Postgres, RocksDB, and Cassandra.
Thompson, Farley, Barker, and co. on the Disruptor: ring buffers, sequence barriers, and avoiding locks when microseconds are the product.
What the Flight protocol ships, why a fat props object becomes a fat HTML-adjacent stream, and how to keep RSC responses small.