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.
Tag
265 articles · 973 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.
Apple's Core ML runtime runs converted models on CPU, GPU, and Neural Engine so features like camera effects and keyboard suggestions do not need a round trip to a datacenter.
Swift's async/await, Task, and actors gave Apple's platforms a structured concurrency model so cancellation and isolation could be compiler-visible instead of GCD folklore.
Apple's Secure Enclave Processor holds biometric templates and private keys in a separated world so iOS can prove 'the user is present' without exposing key material to the application processor.
Apple's CloudKit gives apps a record-based, private-and-public database with serverside sync so iPhone and Mac copies of data converge without each developer inventing iCloud Drive semantics.
YouTube's Content ID system fingerprints audio and video so new uploads can be matched against a rights-holder catalog at a scale where naive hashing would miss remixes and catch nothing.
YouTube Live has to take an unpredictable ingest, transcode in real time, and distribute to millions with a latency budget measured in seconds, not hours.
YouTube turns each upload into a ladder of encoded renditions and lets the player switch rungs as bandwidth changes, which is how video stays watchable on phones and TVs worldwide.
YouTube built Vitess to shard MySQL with a proxy, a lock server, and resharding workflows so a massive video site could keep SQL instead of betting the company on a new store.
NVIDIA's Megatron-LM paper showed tensor and pipeline parallelism so Transformer layers larger than one GPU's memory could still train efficiently on DGX-class boxes.
TensorRT lowers trained networks into fused, precision-calibrated kernels so production inference can use Tensor Cores instead of running training graphs unchanged.
NCCL implements all-reduce and friends across GPUs and nodes so data-parallel training can use NVLink, NVSwitch, and InfiniBand without each framework rewriting rings.
NVIDIA's CUDA turned GPUs from graphics pipelines into a SIMT programming model that scientific computing and later neural networks could actually ship on.
OpenAI released tiktoken so applications could use the same byte-pair encoding as the API, making context limits and billing visible before a request is sent.
OpenAI's Whisper paper turned weakly supervised web audio into a multilingual transcriber; production use then had to solve chunking, latency, and punctuation as a serving problem.
How OpenAI turned a sequence of larger GPT models into a versioned HTTP API with tokens, rate limits, and compatibility constraints that application teams could build on.
How Snowflake scales a virtual warehouse out with extra clusters so queued queries get hardware instead of waiting behind a dashboard storm.
Snowflake stores tables as compressed micro-partitions with per-column min/max metadata so most queries never open most files.
Dageville et al. at SIGMOD 2016 described a warehouse that stores data in object storage and spins independent virtual warehouses for compute.
Databricks open-sourced MLflow so teams could record parameters, metrics, and artifacts from training runs without each group inventing a spreadsheet protocol.
Unity Catalog centralized identities, tables, and permissions so lakehouse objects could be governed like warehouse objects instead of as raw S3 prefixes.
Databricks built Photon as a native, vectorized engine so SQL and Spark workloads could leave interpreted row-at-a-time execution on the table.
Delta Lake put an ordered transaction log in front of Parquet files so Spark jobs could get ACID tables on cheap object stores that do not speak databases.
Zaharia et al.'s NSDI 2012 RDD paper showed that coarse-grained lineage can rebuild lost partitions without replicating every intermediate byte.