Maximum Hashrate.
Zero AI Contention.
Hand-scheduled binary SASS kernels for NVIDIA RTX 30, 40 & 50 Series and Intel Arc Xe2.
Mine at peak silicon efficiency or co-schedule simultaneously alongside local LLMs with 0.0% GDDR bandwidth conflict.
Choose Your Durinn Edition
Every binary ships with pre-compiled multi-architecture .cubin modules (sm_80–sm_120a) and direct NVML hardware clock/V-F control.
durinn-web
Standalone single binary with embedded dark Web-UI, interactive CrashGuard V/F calibration, and 1-click self-updater.
- ✓ Live Web Dashboard on
127.0.0.1:21551 - ✓ Automated Pure & built-in
LLM Synthcalibration - ✓ Built-in Bug & Benchmark uploader
durinn-cli
Ultra-lean headless daemon with UNIX domain socket control, JSON telemetry API, and sub-millisecond AI phase hooks.
- ✓ Zero external runtime dependencies
- ✓ Loopback API on
:21550&durinn.sock - ✓ Instant
POST /api/phaseAI phase switching
Docker & Compose
Lightweight container stack for nvidia-container-toolkit alongside Ollama, vLLM, and llama.cpp.
- ✓ Zero 4 GB CUDA image — mounts host driver
- ✓ Shared
durinn.sockAI Co-Mining sidecar - ✓
SYS_ADMINNVML clock &810 MHzVRAM lock
HiveOS & MMPOS
Custom Miner .tar.gz packages integrating durinn-cli directly into HiveOS Flight Sheets, MMPOS, and msOS.
- ✓ Per-GPU TH/s, Watt, Temp & Fan telemetry
- ✓ Automatic
810 MHzGDDR lock in Pure mode - ✓ Shared Merkle cache for
PCIe x1risers
GPU Profiles & How 4:3 Profiling Works
A single static overclock cannot stay optimal when your GPU shifts between dedicated mining, prompt prefill, and token streaming. Durinn solves this with 4:3 GPU Profiling: 12 pre-verified hardware slots per GPU that adapt automatically in < 1 ms.
Four Live Operating Regimes
Durinn tracks your AI server's exact execution state and applies a dedicated hardware tune for every phase:
4 Phases × 3 Goals = 12 Slots
Every phase includes three distinct operating targets — 🌱 Efficiency (minimum watts & Vmin undervolt), ⚖️ Balanced (sweet-spot TH/s & AI retention), and 🔥 Max Power (peak throughput) — backed by up to 57 silicon-measured tunes (✓ Verified & 🧪 LLM Synth) per GPU.
Zero Manual Tuning Files
Open any GPU model on /gpu-profiles, mix and match any of the 57 hardware tunes across your 4:3 matrix, and paste your generated /p/<gpu> link straight into durinn-web or durinn-cli.
⚡ Quickstart & Copy-Paste Setup
Zero CUDA Toolkit required. Pick your environment below and copy the ready-to-run configuration.
🛡️ Why Durinn Leads on Every GPU
Custom Clocks & V/F Curves Out of the Box
Standard NVIDIA drivers always ship with both libcuda (for loading our pre-built SASS .cubin modules in 0.25s) and libnvidia-ml (NVML). Durinn calls NVML directly from native Go to lock core clocks, lock 810 MHz VRAM in Pure mode, and step V/F undervolt offsets safely with CrashGuard — zero 4 GB CUDA Toolkit needed.
Simultaneous LLM Inference Without Starvation
Traditional miners saturate GDDR6X/GDDR7 memory controllers and crush token generation speed. Durinn synthesizes tile permutations on-chip inside Shared Memory & Tensor Cores (0.0% memory controller load) and adapts GPU scheduling in < 1 ms via /api/phase.
Honest Task-Attributed Watt Accounting
When Mining and an LLM share a GPU, Durinn measures GPU idle base power, solo LLM draw, and combined load — splitting shared idle power 50/50 across both tasks so Mining W + LLM W = Total W, reporting true TH/s/W and tok/W per task.
Unforgeable Proof-of-Benchmark Leaderboard
Leaderboard submissions cannot be spoofed with scripts: the server issues a 90-second random 256-bit challenge seed, your GPU executes real Tensor-Core PearlHash tiles, and the server verifies the 128-bit cryptographic tile signatures mathematically on ingest.
Which GPUs and drivers are supported?
NVIDIA Ampere (RTX 3060–3090 Ti, sm_86), Ada Lovelace (RTX 4060–4090, sm_89), Hopper/Data Center (sm_80/sm_90a), and Blackwell (RTX 5050–5090, sm_120a) on Linux x86_64 with NVIDIA driver 550+ (570+ for RTX 50-series), plus Intel Arc Xe2 Battlemage via Level Zero.
What happens if a V/F undervolt offset is too aggressive for my silicon?
Durinn writes a synchronous CrashGuard breadcrumb to disk before testing any clock or V/F offset in an isolated child process and validates GPU output bit-for-bit against a CPU reference. If a chip rejects an offset, Durinn automatically steps back to a verified stable ceiling.
How do I submit bug reports, feature requests, or link benchmarks to my rig alias?
Click the profile icon in the top-right corner to create an account. Once signed in, the 🐞 Report Bug button appears directly in the header (and inside durinn-web), and your Account page provides your personal token to badge verified GPU benchmarks.