Επίτευξη Εκτέλεσης Πραγματικού Χρόνου σε Πολυμισθωτικές GPUs μέσω Μικροαρχιτεκτονικών Βελτιστοποιήσεων

Abstract

Edge robotics consolidates latency-critical control loops and throughput-oriented generative-AI workloads onto a shared GPU under strict size, weight, power, and cost (SWaP-C) constraints. Because commercial GPUs optimize for aggregate throughput, contention within the shared memory hierarchy destroys the timing guarantees robotic autonomy requires. The prevailing industrial mitigation, static spatial partitioning, isolates compute cores; however, memory injection remains entirely unregulated.

The resulting failures are immediate and measurable. On a physical V100 GPU, co-locating a manipulation policy with a vision-language model inflates 99th-percentile iteration latency by 6.2× and forces the critical task to miss 100% of its safety deadlines. Across four robotics tasks and five co-runners, the critical tail inflates by 2.00× to 49.59×. Cycle-accurate simulation identifies structural memory queuing as the dominant interference driver. Decomposing the cost of spatial partitioning into surrendered compute and residual unmitigated interference exposes the pairings where static separation fundamentally fails to deliver isolation. This thesis targets exactly those scenarios.

We propose two hardware mechanisms that recover deadlines by eliminating memory work, moving beyond simple request reordering. Upstream, a Global CTA Budget caps best-effort threadblock residency device-wide before dispatch, bounding the memory injection rate at its source. It eliminates deadline misses entirely, sustains up to 2.74× higher co-runner throughput than optimal spatial partitioning, and costs 1,692 transistors alongside a two-pass selection policy that prevents dispatcher starvation. Downstream, Dead-Block Bypass targets capacity misallocation at the L2 cache. Nearly half of all best-effort lines are evicted without a single reuse. A 362 KB partial-tag predictor identifies these dead lines and skips their installation. Critical requests remain strictly immune to this bypass. Across 15 pairings it lowers critical tail latency without conceding best-effort throughput in 11 cases, and raises best-effort throughput by up to 46.4% against generative-AI decode workloads.

Recovering strict temporal deadlines therefore requires reducing the absolute volume of work the shared structures must absorb. Downstream reordering alone proves structurally insufficient for this task. Upstream concurrency control, downstream traffic elimination, and spatial partitioning each address distinct operational regimes, separated by boundaries that are entirely measurable prior to deployment.