An engineering-oriented comparison of KAI-Scheduler’s Reservation Pod approach and HAMi’s hard isolation path, including trade-offs, failure modes (noisy neighbor), and how the two layers can complement each other.
Category archive
GPU Acceleration
Posts under GPU / GPU Acceleration.
GPU / GPU Acceleration collects 3 posts focused on practical patterns, operations context, and implementation details so readers can move from concepts to production decisions.
More topics
Other topics
Selected pieces
Featured in this category
An engineering-oriented guide to hetGPU: how a compiler + runtime stack can make one GPU binary run across NVIDIA/AMD/Intel/Tenstorrent, including SIMT vs MIMD, memory model gaps, and live kernel migration.
A deep dive into Kubernetes GPU virtualization through gpu-manager startup flow, including device interception, topology awareness, scheduling, and allocation mechanics.
Archive
All posts in this category
All posts in reverse chronological order.
KAI-Scheduler vs HAMi: Two Ways to Share GPUs in Kubernetes (Soft vs Hard Isolation)
An engineering-oriented comparison of KAI-Scheduler’s Reservation Pod approach and HAMi’s hard isolation path, including trade-offs, failure modes (noisy neighbor), and how the two layers can complement each other.
hetGPU: Chasing Cross-Vendor GPU Binary Compatibility
An engineering-oriented guide to hetGPU: how a compiler + runtime stack can make one GPU binary run across NVIDIA/AMD/Intel/Tenstorrent, including SIMT vs MIMD, memory model gaps, and live kernel migration.
Kubernetes GPU Virtualization Explained Through gpu-manager Startup Flow
A deep dive into Kubernetes GPU virtualization through gpu-manager startup flow, including device interception, topology awareness, scheduling, and allocation mechanics.