jamesob
Featured
local-llm
A comprehensive guide for building and configuring a high-end local machine to run state-of-the-art LLMs, with detailed hardware choices, BIOS tuning, and Docker-based model serving.
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It provides a rigorous, evidence-based methodology to separate 'it ran' from 'it is safe', including a dedicated gate that catches silent corruption at specific undervolt offsets, and tools for soak testing under real inference workloads.
Who it is for
- Owners of NVIDIA CMP 170HX (GA100) GPUs with the 64 GB memory unlock
- AI inference operators seeking to reduce power without sacrificing correctness
- Overclockers and miners who need stable undervolt profiles for long-running loads
- Developers working on GPU reliability tooling or voltage/frequency characterization
Use cases
- Undervolting a CMP 170HX to lower power draw by ~29% while maintaining stock throughput
- Running a multi-sweep pattern gate to verify memory integrity before shipping a profile
- Soaking a candidate undervolt with a real inference workload (e.g., vLLM, sglang) to catch intermittent faults
- Recovering from a device fault or hang using a recover command that also checks for context-creation wedge
Strengths
- Detects silent memory corruption that passes other stability tests (e.g., +325/1400 passes but corrupts memory)
- Provides concrete benchmark numbers (GEMM, serving throughput, power) with per-card variability acknowledged
- Includes a soak tool (170hx-soak) that repeats workloads and monitors Xid errors, catching faults missed by shorter tests
- Robust safety measures: armed marker, boot-check, quarantine system, and a context probe that identifies wedges invisible to nvidia-smi
Considerations
- Specific to the NVIDIA CMP 170HX with driver 610.43.03 and the community memory unlock; other GPUs are refused
- Undervolt offsets are per-card and must be individually qualified; results from the reference card may not transfer
- Requires root access, bash, NVML, and a CUDA toolkit; the setup process is non-trivial and assumes a patched driver environment
README quick start
Quick start
Description
Tuning and qualification harness for the NVIDIA CMP 170HX (GA100). Measures, gates and recovers clock/voltage settings, and treats a completed benchmark as evidence of nothing.
Related repositories
Similar projects matched by category, topics, and programming language.
bjan
Featured
claudemux
Claudemux is a lightweight, dependency-minimal message bus that lets multiple Claude Code sessions communicate in real-time over tmux, enabling one session to ask another a question and receive the answer without manual intervention.
AARomanov1985
Audio-Cassette-Simulation
A set of bash scripts using ffmpeg to simulate vintage cassette tape audio profiles with noise, wow/flutter, bandwidth limits, and EQ.