Скачать или смотреть Moving Past Speculation: How Deterministic CPUs Deliver Predictable AI Performance

For decades, modern Central Processing Units (CPUs) have relied on *speculative execution* to maintain speed by predicting instruction outcomes and filling execution pipelines. While effective, this approach introduced significant drawbacks, including wasted energy when predictions failed, increased design complexity, and exposure to security flaws like Meltdown and Spectre. A new alternative execution model, documented in a series of recently issued patents, offers a fundamentally *deterministic, time-based architecture* that challenges this standard paradigm. Instead of relying on guesswork, this model uses a simple, cycle-accurate *time counter* and a Register Scoreboard to assign a precise execution slot for each instruction only when its data dependencies and necessary resources are confirmed to be ready. This deterministic scheduling ensures a predictable, pre-planned flow of execution, maintaining the efficiency of out-of-order execution while eliminating the unpredictable performance cliffs and power waste caused by failed speculative rollbacks. By simplifying hardware and ensuring high utilization of execution units without requiring complex mechanisms like register renaming, this time-based design is particularly well-suited to the high-throughput vector and matrix computations that dominate modern Artificial Intelligence (AI) and Machine Learning (ML) workloads.

https://venturebeat.com/ai/moving-pas...