Dedicated acceptance test suite for NVIDIA GeForce RTX 3090 cards modified from 24GB to 48GB VRAM.
Test. Validate. Reject bad mods before handover.
--- ## Why This Exists RTX 3090 48GB mods are not normal GPUs. A card can boot, display output, and still fail handover because the driver reports the wrong memory size, CUDA sees less VRAM than expected, high-memory allocation crashes, or the OS logs NVIDIA, WHEA, Xid, or PCIe errors. This repo is a workshop acceptance kit. Run it before accepting the card and keep the generated report as evidence. ## 3090 48GB-Specific Checks - GPU name must contain `3090` - VRAM must report about `49152 MiB` - PyTorch CUDA must see about the same 48GB memory - CUDA compute capability should be RTX 3090-class `8.6` - VRAM allocation writes to 92 percent of the 48GB target by default - PCIe link, VBIOS, driver, power, temperature, and utilization are captured - Linux NVIDIA/Xid/PCIe kernel errors are checked - Windows NVIDIA/Display/WHEA/PCIe System events are checked - Every report gets an `ACCEPT`, `REVIEW`, or `REJECT` verdict ## Quick Start ### Linux ```bash chmod +x install.sh gpu-validator.sh ./install.sh ./gpu-validator.sh --quick ./gpu-validator.sh ``` ### Windows PowerShell ```powershell Set-ExecutionPolicy -Scope Process Bypass .\install.ps1 .\gpu-validator.ps1 --quick .\gpu-validator.ps1 ``` ### Windows Command Prompt ```bat gpu-validator.bat --quick gpu-validator.bat ``` Reports are saved in `reports/` as Markdown and JSON. ## Recommended Workshop Run Run the quick check first: ```bash ./gpu-validator.sh --quick ``` Then run a stricter handover test: ```bash ./gpu-validator.sh --stress-seconds 1800 --vram-fraction 0.95 ``` On Windows, use the same options with `.\gpu-validator.ps1`. Before leaving the workshop, reboot once and run `--quick` again. ## Common Options ```bash ./gpu-validator.sh --gpu-index 0 ./gpu-validator.sh --quick ./gpu-validator.sh --stress-seconds 900 ./gpu-validator.sh --vram-fraction 0.95 ./gpu-validator.sh --skip-stress ./gpu-validator.sh --expected-memory-mib 49152 --memory-tolerance-mib 1024 ``` ## Verdicts | Verdict | Meaning | | --- | --- | | `ACCEPT` | No failed or warning checks | | `REVIEW` | No hard failures, but something needs human review | | `REJECT` | One or more RTX 3090 48GB acceptance checks failed | Do not accept the card on `REJECT`. Treat `REVIEW` as pause-and-investigate, especially for PCIe warnings, OS log warnings, temperature issues, or missing kernel/event-log permissions. ## Acceptance Checklist - `nvidia-smi` selected GPU name contains `3090` - `nvidia-smi` reports about `49152 MiB` VRAM - PyTorch CUDA reports about the same memory - CUDA compute capability is `8.6` - VRAM allocation reaches the configured 48GB target fraction - Matrix stress completes without driver reset, crash, artifacting, or reboot - Linux report has no fresh `NVRM`, `Xid`, or PCIe errors - Windows report has no NVIDIA/Display/WHEA/PCIe warning or error events - Temperature and power remain reasonable for the cooler and PSU - Quick test still passes after a reboot ## Files | File | Purpose | | --- | --- | | `gpu_validation_suite.py` | Main validator | | `gpu-validator.sh` | Linux runner | | `install.sh` | Linux installer | | `gpu-validator.ps1` | Windows PowerShell runner | | `install.ps1` | Windows installer | | `gpu-validator.bat` | Windows Command Prompt wrapper | | `other-commands.txt` | Manual workshop commands | ## Important Limits This suite cannot prove long-term reliability or soldering quality. Also inspect memory IC markings, thermal pad contact, cooler mounting, PCB work, VBIOS state, fan behavior, cold boot behavior, and stability in the actual workloads you care about.