Barswap
GPU multiplexer for dense compute systems
Unlock all your GPU hardware on commodity motherboards. Run inference and training across every GPU die you own, regardless of system address space limits.
The Problem & The Solution
Budget motherboards can't address all your GPU hardware. Barswap fixes that.
The Problem
Multi-GPU cards pack multiple dies per board, but commodity motherboards don't have enough address space to map them all. The OS sees the hardware, but can't use it. Most of your GPU dies sit completely dark.
You bought 72 GB of VRAM. You can use a fraction of it. The hardware works fine — the system just can't reach it.
The Solution
Barswap is a Linux kernel module that dynamically manages GPU access, letting the system use every die across all your cards. Model weights and compute state persist between switches — no data lost, no reload needed.
Drop it in, load the module, and your full GPU cluster comes online.
Status & Roadmap
From kernel module to full GPU cluster.
Core GPU Multiplexer
Kernel module working. All GPU dies accessible and validated. Full CUDA compatibility.
GPU-Accelerated Inference
Running language models on multiplexed GPUs with significant speedup over CPU-only inference. Output validated bitwise-identical.
Inference Engine Integration
Integrating with our inference engine for seamless multi-model serving. Expanding the GPU cluster with additional cards.
Multi-GPU Concurrent Inference
Serving multiple models simultaneously across all GPU dies with automatic scheduling.
Training Support
Fine-tuning model adapters across multiplexed GPU dies.
Public Release
Packaged release with installer and configuration tools for supported hardware setups.
The Stack
Barswap is one layer in a fully integrated, zero-dependency AI infrastructure.
ML Framework
Compute primitives, quantization, training.
Barswap
GPU multiplexing. Hardware layer.
Inference Engine
Model serving. Text, audio, vision.
TeamIDE
Desktop IDE with built-in local AI.
Barswap is the hardware layer in a vertically integrated stack. From compute primitives to model serving to the IDE — every layer is built in-house with zero external dependencies.
Stay Updated
This is a solo project. Your support helps cover hardware costs and keeps development going.