Heterogeneous GPU Sharing on Kubernetes

KVM Backend for VirtualBox

A transparent, in-container GPU resource controller that enforces memory and compute limits by intercepting CUDA calls without application or driver changes.

Here are my personal paper reading notes (including machine learning systems, AI infrastructure, and other interesting stuffs).

Tensor Fusion is a state-of-the-art GPU virtualization and pooling solution designed to optimize GPU cluster utilization to its fullest potential.

KPilot: Unified control plane for multi-cluster Kubernetes management, GPU compute scheduling, and model serving.

Research: vGPU unlock on consumer NVIDIA RTX 5090 (Blackwell/GB202). 19 binary patches, full CPU-side pipeline working, GSP firmware blocked by fused-off VF PRIV registers.

A GPU Virtualisation benchmarking tool - For LLM like workloads, overhead evaluations and on resource isolation metrics

TensorFusion landing page and product docs

Transparent library for CUDA/NVML that virtualizes GPU access, enforces memory limits, and enables container-friendly sharing with minimal performance overhead.

Bitfusion (bitfusion.io) was an Austin, Texas startup, backed by DCVC, that built "elastic AI" infrastructure software. Its FlexDirect platform was a transparent GPU/CPU virtualization layer that pooled network-attached accelerators (GPUs, FPGAs, ASICs) across TCP, RoCE, and InfiniBand into a single elastic compute cluster for sharing, scaling…

🎮 Enable GPU passthrough for KVM on full AMD systems, simplifying setup for Arch Linux and Windows 11 virtual machines.

Bitfusion (Bitfusion.io) was a GPU virtualization startup whose FlexDirect software pooled and disaggregated GPUs across a network so machine learning and AI workloads could share accelerator hardware on demand.

HAMI (Heterogeneous AI Computing Virtualization Middleware) is an open-source project that enables GPU virtualization and sharing for Kubernetes workloads, allowing multiple AI containers to share GPU resources efficiently.