Skip to content

unified_memory.md

Latest commit

 

History

History
80 lines (61 loc) · 2.22 KB

unified_memory.md

File metadata and controls

80 lines (61 loc) · 2.22 KB

Unified Memory Guide

Overview

Unified Memory allows GPUs to access system RAM when GPU VRAM is insufficient. Essential for running large structures like 6QNR on consumer GPUs.

When to Enable

Enable Unified Memory:

  • Large structures exceeding GPU memory
    • 6QNR requires ~15GB, RTX 4080 has 16GB total
  • Getting out-of-memory errors
  • Consumer GPUs with limited VRAM
    • RTX 4080 (16GB) + 6QNR ✅
    • RTX 3080 (10GB) + large proteins ✅

Don't Enable:

  • High-end GPUs with sufficient VRAM
    • H100 (80GB) ❌
    • A100 (40GB) ❌
    • RTX 4090 (24GB) ❌

Configuration

Edit your config file:

nano my_system.config

# Change this line:
UNIFIED_MEMORY=true

How It Works

When enabled, AFSysBench automatically sets:

XLA_PYTHON_CLIENT_PREALLOCATE=false    # Don't pre-allocate GPU memory
TF_FORCE_UNIFIED_MEMORY=true           # Enable TensorFlow unified memory
XLA_CLIENT_MEM_FRACTION=3.2            # Request 320% of GPU memory

Performance Impact

Configuration 6QNR Runtime Notes
H100 (80GB) ~5 minutes No unified memory needed
RTX 4080 + Unified ~8-10 minutes 60-100% slower but works
RTX 4080 without Fails Out of memory

Example: 6QNR on RTX 4080

# 1. Enable unified memory
nano my_system.config
# Set: UNIFIED_MEMORY=true

# 2. Run benchmark
python af_bench_runner_updated.py -c my_system.config inference -i 6QNR_subset_data.json -t 1

# 3. Expected output in log:
# Running model inference with seed 1 took 486.67 seconds
# Fold job 6QNR_subset done, output written to /output/6qnr_subset

Success Indicators

✅ Log shows: "Running model inference with seed 1 took XXX seconds"
✅ Output contains: *.cif, *_confidences.json files
✅ GPU memory stays at ~97% (not crashing)
✅ System RAM usage increases during run

Troubleshooting

Still Getting OOM Errors

  1. Verify: grep UNIFIED_MEMORY my_system.config shows true
  2. Check: Sufficient system RAM available (free -h)
  3. Ensure: No other GPU processes running

Too Slow Performance

  1. Close other GPU applications
  2. Ensure fast system RAM (DDR4-3200+ recommended)
  3. Consider upgrading to 24GB+ GPU for production use