A practical guide to GPU overprovisioning strategies, including scheduler-level oversubscription, time slicing, memory controls, MIG, vGPU, queue backfill, and operational guardrails.
Category archive
GPU
Posts under GPU.
GPU collects 5 posts focused on practical patterns, operations context, and implementation details so readers can move from concepts to production decisions.
More to explore
Explore more topics
Selected pieces
Featured in this category
A practical guide to choosing between serverless GPUs and dedicated GPUs for startups, based on cost structure, delivery speed, performance predictability, operations burden, and team maturity.
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.
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 stream
All posts in this category
Browse the full archive in reverse chronological order.
GPU Overprovisioning Solutions: From Oversubscription and Sharing to Isolation
A practical guide to GPU overprovisioning strategies, including scheduler-level oversubscription, time slicing, memory controls, MIG, vGPU, queue backfill, and operational guardrails.
How Startups Should Choose: Serverless GPU vs Dedicated GPU
A practical guide to choosing between serverless GPUs and dedicated GPUs for startups, based on cost structure, delivery speed, performance predictability, operations burden, and team maturity.
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.