Releases·Tool Release
Show HN: CUDA/graphics in QEMU-KVM VMs without passing the Nvidia card to them
A developer released an open-source tool that runs CUDA and Vulkan graphics inside QEMU-KVM VMs on a shared GeForce card by forwarding the NVIDIA driver's ioctl surface from guest to host, avoiding VFIO/vGPU passthrough. Guests run stock libcuda and a real Vulkan ICD with no API remoting or translation, while the host desktop keeps using the same GPU. This matters to practitioners who need GPU acceleration in VMs without dedicating a physical card.
7 engagement1 sourceTue, Sep 8, 2026, 12:25 PM
AI Overview
- GitHub project forwards NVIDIA driver ioctl surface from guest to host, letting stock libcuda and a real Vulkan ICD run in the guest
- Tool supports CUDA, Vulkan graphics, display, LLM, and headless workloads with no VFIO/vGPU passthrough or API remoting
- The host desktop keeps using the same GeForce card while multiple VMs share it
Related
Products·launched
Infinigpu: GPU sharing across VMs via vibecoding
A proof-of-concept GPU splitter that allows a single GPU to be shared by multiple virtual machines. Built in two days using Claude Code Opus 4.8, it demonstrates rapid prototyping of low-level kernel software through LLM-guided development, though it currently has many bugs, input lag, and no sound.
A proof-of-concept GPU splitter that allows a single GPU to be shared by multiple virtual machines. Built in two days using Claude Code Opus 4.8, it demonstrates rapid prototyping of low-level kernel software through LLM-guided development, though it currently has many bugs, input lag, and no sound.
2 engagement1 sourceredditFri, Jul 17, 2026, 11:59 PM
Social·Community
Users report Ollama detecting RTX 5090 as Vulkan device in dual-GPU setups
A user reports that Ollama detects their RTX 5090 as a Vulkan device instead of CUDA in a dual-GPU setup, despite setting CUDA_VISIBLE_DEVICES. The issue appears related to environment variable handling, potentially affecting performance for Qwen 3.8 inference.
A user reports that Ollama detects their RTX 5090 as a Vulkan device instead of CUDA in a dual-GPU setup, despite setting CUDA_VISIBLE_DEVICES. The issue appears related to environment variable handling, potentially affecting performance for Qwen 3.8 inference.
0 engagement1 sourceredditSat, Aug 15, 2026, 12:21 PM
Social·Community
Users seek ways to utilize idle VRAM across L40 GPUs in vLLM clusters
An administrator managing five Proxmox hosts with dual NVIDIA L40 GPUs reports that vLLM-served LLMs leave about 5 GB of VRAM idle per card despite 90% utilization. They ask for solutions to pool or use the free VRAM across cards, given heterogeneous model sizes.
An administrator managing five Proxmox hosts with dual NVIDIA L40 GPUs reports that vLLM-served LLMs leave about 5 GB of VRAM idle per card despite 90% utilization. They ask for solutions to pool or use the free VRAM across cards, given heterogeneous model sizes.
5 engagement1 sourceredditTue, Aug 18, 2026, 10:20 AM
Social·Community
User explores Vulkan-based cross-vendor GPU inference pooling across AMD and NVIDIA
A Reddit user with two AMD GPUs (R9700, 9070XT) and two NVIDIA RTX 3060 12GB GPUs on separate machines proposes using Vulkan to pool them into a single VRAM resource for inference. The idea is experimental and likely slow, but could enable larger model inference on heterogeneous hardware.
A Reddit user with two AMD GPUs (R9700, 9070XT) and two NVIDIA RTX 3060 12GB GPUs on separate machines proposes using Vulkan to pool them into a single VRAM resource for inference. The idea is experimental and likely slow, but could enable larger model inference on heterogeneous hardware.
18 engagement1 sourceredditThu, Jul 23, 2026, 12:18 AM
Social·Community
Developers weigh 64–128GB unified memory vs CUDA for €6k AI laptop
A developer asks whether a €6k portable AI/development setup should prioritize 64–128GB unified memory (M5 Pro/Max MacBook Pro) or CUDA compatibility (high-end NVIDIA laptop). The trade-off centers on local inference and ML experiments versus GPU memory, heat, noise, and battery life.
A developer asks whether a €6k portable AI/development setup should prioritize 64–128GB unified memory (M5 Pro/Max MacBook Pro) or CUDA compatibility (high-end NVIDIA laptop). The trade-off centers on local inference and ML experiments versus GPU memory, heat, noise, and battery life.
12 engagement1 sourceredditTue, Aug 25, 2026, 11:59 AM