🚀 Welcome to RamShared v0.12.0 — Using Idle GPU VRAM as Ultra-Fast RAM Cache for Linux & WSL2

[emersonbusson]

Hey everyone! 👋 Welcome to RamShared.

If you've ever compiled a large Rust, C++, or Go project, trained a local AI model, or spun up heavy Docker containers inside WSL2 or Linux, you've likely experienced the infamous desktop freeze.

When memory fills up, standard operating systems swap memory pages to disk. On Windows/WSL2, swap must traverse four nested virtualization layers (ext4 ➔ VHDX ➔ Hyper-V ➔ NTFS). Under heavy load, disk writes crawl down to 60–80 MB/s, locking up the entire system for 30–60 seconds.

Meanwhile, your GPU (NVIDIA, AMD, or Intel) is often sitting right there with gigabytes of high-bandwidth VRAM doing absolutely nothing.

---

💡 The Core Idea: Borrowing Idle VRAM as an Ultra-Fast Memory Tier

RamShared routes overflow memory pages directly across high-speed PCIe bus lines into your idle GPU memory (reaching 6.0+ GB/s sustained bandwidth), keeping Linux and WSL2 silky smooth and completely responsive.

🛡️ Games and AI Workloads Always Win (Zero Risk)

VRAM is leased purely as a smart, temporary cache. All writes are backed by persistent storage origin. The instant a game, 3D application, or local AI workload (such as PyTorch, Ollama, or Stable Diffusion) requests GPU memory, RamShared yields the VRAM back in milliseconds without dropping applications or losing a single byte of data.

⚡ 100% Optional GPU

Don't have a dedicated GPU? Headless server? Running on an integrated chip? RamShared works seamlessly without a GPU too, orchestrating fast in-RAM compression (ZRAM) and SSD storage with the exact same freeze-proof stability.

---

🚀 What's New & Qualified in Release v0.12.0

We just published Release v0.12.0 with major stability and architectural advancements:

1. 4 GiB Active VRAM Tiering: Fully qualified on dedicated physical hardware (NVIDIA GeForce RTX 2060).
2. Dynamic Host Display Safety Floor: Enforced dynamic headroom clamping (max(1536 MB, 20%)) ensuring the Windows host desktop and Desktop Window Manager (DWM) never stutter or drop frames.
3. Closed-Loop Swap Governor: Paces memory allocation against active kernel kswapd drain rates, eliminating tail latency spikes.
4. 171% RAM Stress Qualification (PASS_ZERO_PANIC): Sustained 20,208 MB active memory pressure across RAM, ZRAM, 4 GiB VRAM, and SSD swap with zero OS stalls, 100% SHA-256 byte-exact data integrity, and 21.66 GB/s DMA reclaim in 910 ms.
5. Empirical Telemetry: Integrated real-time /proc/diskstats throughput tracking (showing 31.2x acceleration over SSD fallback) and p99 tail jitter metrics into the interactive TUI (ramshared top).
6. Pre-Built Linux Packages: Automated .deb (Ubuntu/Debian), .rpm (Fedora/RHEL), and Arch Linux .tar.gz packages published directly on GitHub Releases.
7. Linux 6.18 LTS Kernel Upstreaming: In-tree C block driver architecture and upstream RFC submitted to Microsoft WSL and LKML (microsoft/WSL #41054).

---

📊 Hardware Benchmark Snapshot (Physical Silicon)

Category / Metric	Tier 0: Compressed RAM (ZRAM)	Tier 1: GPU VRAM (PCIe DMA)	Tier 3: SSD Storage Origin	Verdict
Access Latency	0.08 µs	1.72 µs	48.2 µs	🟢 31.2x faster than SSD
Throughput	Direct CPU RAM	6.07 GiB/s PCIe	6.63 GB/s raw	🟢 Ultra-high bandwidth
Empirical Throughput	128.2 MB/s	625.4 MB/s	1,077.2 MB/s	🟢 Accelerated swap path
Active Saturation	1,024 MB (100%)	4,096 MB (100%)	2,367 MB active	🟢 171% of total RAM
Reclaim Bus Bandwidth	Direct RAM	21.66 GB/s DMA	Direct Disk	🟢 Full reclaim in 910 ms
PSI Memory Stall Time	0.00% avg10	0.00% avg10	0.00% avg10	🟢 Zero system stalls
Bit-Exact Integrity	100% Match	100% Match	100% Match	🟢 Zero bit flips
Stability Verdict	PASS	PASS	PASS	🟢 PASS_ZERO_PANIC

---

🛠️ Quick Start

# 1. Build and install release binaries
./scripts/quickstart.sh

# 2. Check your environment and GPU topology
ramshared check

# 3. Launch the interactive real-time dashboard
ramshared top

Pre-compiled binary packages (.deb for Ubuntu/Debian, .rpm for Fedora/RHEL/CentOS, and Arch Linux tarballs) are available on the Releases page.

---

💬 Community & Feedback

We want to hear from you:

• What GPU / distro hardware configuration are you testing on?
• What are your thoughts on direct in-tree kernel drivers (ublk / io_uring) vs user-space daemons?
• Have suggestions for multi-vendor support (AMD ROCm, Intel Arc OneAPI, Vulkan VMA)?

Drop your questions, benchmarks, and ideas below! All tests and discussions are welcome. 🙌