Scruff AI
159 lines
4.6 KiB
Markdown
159 lines
4.6 KiB
Markdown
# GTX 1050 Deployment & Testing Plan
|
||
|
||
## 🎯 **Target Hardware:**
|
||
- **OS:** Ubuntu 24.04 LTS
|
||
- **CPU:** Intel i7-7700HQ @ 2.80GHz (4 cores, 8 threads)
|
||
- **RAM:** 32GB
|
||
- **GPU:** NVIDIA GTX 1050 4GB (nvidia-driver-470)
|
||
- **Disk:** 937GB NVMe (~87GB used, ~803GB free / 10%)
|
||
|
||
## 📦 **What We Know Works:**
|
||
1. **256×256 grid** - Compiled and tested on Windows/RTX 4090
|
||
2. **Guardian formation** - 13 guardians with RHO_THRESH=1.00022
|
||
3. **Probe sequence** - A, B, C, D stress tests defined
|
||
4. **Power scaling** - ~37W on RTX 4090 (expect ~40-60W on GTX 1050)
|
||
|
||
## 🚀 **Deployment Steps:**
|
||
|
||
### Phase 1: Environment Setup (Ubuntu)
|
||
```bash
|
||
# 1. Verify CUDA installation
|
||
nvidia-smi
|
||
nvcc --version
|
||
|
||
# 2. Install required libraries
|
||
sudo apt-get update
|
||
sudo apt-get install -y build-essential libnvml-dev
|
||
|
||
# 3. Verify GPU architecture support
|
||
# GTX 1050 = Pascal = sm_61
|
||
```
|
||
|
||
### Phase 2: Transfer & Compile
|
||
```bash
|
||
# 1. Copy source files to Ubuntu
|
||
scp probe_256.cu user@gtx1050:~/fractal/
|
||
scp fractal_habit_256_full.cu user@gtx1050:~/fractal/
|
||
|
||
# 2. Compile on target hardware
|
||
cd ~/fractal
|
||
nvcc -O3 -arch=sm_61 -o probe_256_gtx1050 probe_256.cu -lnvml
|
||
nvcc -O3 -arch=sm_61 -o fractal_habit_256 fractal_habit_256_full.cu -lnvml -lcufft
|
||
```
|
||
|
||
### Phase 3: Initial Test
|
||
```bash
|
||
# 1. Test basic execution
|
||
./fractal_habit_256
|
||
|
||
# 2. Check power usage
|
||
sudo nvidia-smi -pl 60 # Set power limit to 60W
|
||
./probe_256_gtx1050
|
||
|
||
# 3. Monitor with nvidia-smi
|
||
watch -n 1 nvidia-smi
|
||
```
|
||
|
||
## 🔬 **Testing Protocol:**
|
||
|
||
### Test 1: Basic Functionality
|
||
- Run `fractal_habit_256` for 100k steps
|
||
- Verify: Guardian formation (13 guardians)
|
||
- Monitor: Power draw, temperature, stability
|
||
|
||
### Test 2: Power Limiting
|
||
```bash
|
||
# Test different power limits
|
||
sudo nvidia-smi -pl 40 # Minimum sustainable
|
||
sudo nvidia-smi -pl 50 # Balanced
|
||
sudo nvidia-smi -pl 60 # Performance
|
||
sudo nvidia-smi -pl 75 # Max (default)
|
||
```
|
||
|
||
### Test 3: Full Probe Sequence
|
||
- Run `probe_256_gtx1050` with monitoring
|
||
- Focus on crash at cycle ~1112 (Probe C - VRM Silence)
|
||
- Collect complete data for all probe phases
|
||
|
||
### Test 4: Long-term Stability
|
||
- Run for extended period (10,000+ cycles)
|
||
- Monitor for memory leaks, GPU errors
|
||
- Check thermal throttling
|
||
|
||
## 📊 **Data Collection:**
|
||
|
||
### Essential Metrics:
|
||
1. **Power:** Watts (nvidia-smi)
|
||
2. **Temperature:** GPU core temp
|
||
3. **Performance:** Cycles per second
|
||
4. **Stability:** Guardian count, omega values
|
||
5. **Memory:** GPU memory usage
|
||
|
||
### Monitoring Script:
|
||
```bash
|
||
#!/bin/bash
|
||
# monitor_gtx1050.sh
|
||
while true; do
|
||
nvidia-smi --query-gpu=power.draw,temperature.gpu,utilization.gpu,memory.used --format=csv
|
||
sleep 1
|
||
done
|
||
```
|
||
|
||
## ⚠️ **Potential Issues & Solutions:**
|
||
|
||
### Issue 1: CUDA Compatibility
|
||
- **Check:** GTX 1050 = sm_61 architecture
|
||
- **Fix:** Compile with `-arch=sm_61`
|
||
|
||
### Issue 2: Power Limiting
|
||
- **Check:** GTX 1050 power limits (40-75W)
|
||
- **Fix:** Use `nvidia-smi -pl` to set limits
|
||
|
||
### Issue 3: Memory Constraints
|
||
- **Check:** 4GB VRAM usage
|
||
- **Fix:** Monitor with `nvidia-smi --query-gpu=memory.used`
|
||
|
||
### Issue 4: Thermal Throttling
|
||
- **Check:** Temperature > 80°C
|
||
- **Fix:** Improve cooling, reduce power limit
|
||
|
||
## 🎯 **Success Criteria:**
|
||
|
||
### Minimum Viable Product:
|
||
1. ✅ 256×256 grid runs on GTX 1050
|
||
2. ✅ 13 guardians form and persist
|
||
3. ✅ Power draw < 60W sustained
|
||
4. ✅ Temperature < 80°C
|
||
5. ✅ No crashes in first 1000 cycles
|
||
|
||
### Extended Goals:
|
||
1. ✅ Complete probe sequence (A-D) without crash
|
||
2. ✅ Stable operation for 10,000+ cycles
|
||
3. ✅ Power efficiency optimization
|
||
4. ✅ Documentation of performance characteristics
|
||
|
||
## 📋 **Immediate Action Items:**
|
||
|
||
1. **Transfer files** to Ubuntu system
|
||
2. **Compile** with correct architecture (sm_61)
|
||
3. **Set power limit** to 60W for testing
|
||
4. **Run basic test** - verify guardian formation
|
||
5. **Execute full probe sequence** - monitor for crash at cycle ~1112
|
||
|
||
## 🕒 **Time Estimate:**
|
||
- Setup: 30 minutes
|
||
- Compilation: 10 minutes
|
||
- Basic test: 15 minutes
|
||
- Full probe sequence: 30-60 minutes
|
||
- **Total:** 1.5-2 hours
|
||
|
||
## 🎪 **Next Steps After Successful Deployment:**
|
||
|
||
1. **Performance optimization** - tune parameters for GTX 1050
|
||
2. **Extended testing** - 24-hour stability run
|
||
3. **Documentation** - create GTX 1050 performance profile
|
||
4. **Scaling tests** - try 384×384 if 256×256 is stable
|
||
5. **Application development** - build on stable foundation
|
||
|
||
---
|
||
**Key Insight:** The forensic audit showed the system works correctly but has power scaling inefficiencies. On GTX 1050, we're targeting the actual hardware constraints (40-60W), so these "inefficiencies" may actually be acceptable or even optimal for this hardware class. |