PNY LAB ON DEMAND What you can test*

• Training Performance

  Training
  Performance

  • Train a small model to measure real training speed
  • Use known models such as LLaMA 7B or standard vision models
  • Measure examples per second
  • Monitor GPU utilization during training
  • Compare performance with your current hardware

  
• Large-Model Inference

  Large-Model
  Inference

  • Load large models such as Mixtral or LLaMA 70B
  • Measure tokens per second and latency
  • Test how many concurrent requests can be processed
  • Verify memory stability under sustained load

  
• Multi-GPU Scalability

  Multi-GPU
  Scalability

  • Test how performance increases when adding GPUs
  • Check the quality of GPU-to-GPU communication
  • Compare the same job on 1 GPU, 2 GPUs, 4 GPUs, and more

  
• Data Pipeline Performance

  Data Pipeline
  Performance

  • Measure file loading speed from storage
  • Evaluate CPU preprocessing speed
  • Detect whether GPUs are waiting for data
  • Identify bottlenecks in the end-to-end pipeline

  
• GPU Memory & Load Stress Testing

  GPU Memory & Load
  Stress Testing

  • Run long-sequence LLMs (from 8k to 64k tokens)
  • Test very large batch sizes
  • Run heavy tensor operations
  • Observe memory stability, fragmentation, and limits