MIG

In Miyabi-G equipped with NVIDIA GH200 Grace Hopper Superchip, a function called MIG (Multi-Instance GPU), which divides GPU resources (SM (compute cores) and memory) into multiple instances, is enabled in some queues. For more information on MIG, please refer to the "NVIDIA Multi-Instance GPU" (external website).

Anticipated Usage Scenarios

• Cannot use up all the computing resources of one NVIDIA GH200 Grace Hopper Superchip
  → The amount of tokens consumed can be reduced by using MIG use queues.
• Debugging, operational testing, and functional testing of code being implemented and developed
  → Reduce waiting time after job submission by avoiding crowded queues by also using MIG use queues.
• Educational use such as lectures at universities
  → Simultaneous execution of jobs by a large number of participants is possible by also using MIG use queues.

Usage of MIG in Miyabi-G

The MIG use queues in Miyabi-G will be provided with NVIDIA GH200 Grace Hopper Superchip divided into 4 parts. The resources allocated to each are shown in the table below. There is no job submission option to avoid MIG#3 which has low GPU resources (SMs).

MIG Use Queues

Interactive Job Queues

• Tokens are not consumed.

Batch Job Queues

Calculation of Tokens

Tokens are consumed for the execution of each batch job. The token consumption coefficient in MIG use queue is 0.25 per MIG instance. Please refer to the calculation formula and specific examples of calculation shown below.