← Back to the GPU series: all 11 demos

GPU Sharing Lab

One H100, four ways to share it · whole GPU, time-slicing, MPS, MIG

Four teams want the same 80 GB H100: a chatbot, an embeddings batch job, a data scientist's notebook and, later, a fine-tuning job. Pick a sharing mode, then use the buttons on the right to add a noisy neighbour or crash one tenant. The numbers are illustrative, but the behaviour of each mode is how the real thing works.

1 · Pick how the GPU is shared

NVIDIA H100 80GB
Compute
Memory · 80 GB
What Kubernetes sees

    

    

      

        
2 · Stir things up

        

      

      

        
Tenants

        

      

      

        
What happened

        

      

    

  


  

    
The four modes side by side

    

    
MIG needs a data-centre GPU that supports it, such as A100, H100, H200 or B200. Time-slicing and MPS work on most NVIDIA GPUs. MIG can also be combined with time-slicing inside each slice.