What is Power Usage Effectiveness (PUE)?
PUE tells you how much extra energy a data center needs beyond the IT hardware. It’s the quickest way to judge whether a facility is efficient or wasting power
PUE = Total Facility Energy / IT Energy
- 1.2 PUE → Extremely efficient
- 2.0 PUE → Twice as much energy overhead
Lower PUE = greener + cheaper.
MIG (Multi‑Instance GPU) slicing is just a logical partitioning of a single GPU into 7 seven separate GPU Instances. Each GPU Instance behaves like a separate GPU for CUDA applications.
This is done as a GPU pass-through virtualization to Linux guests on top of supported hypervisors or via Bare-metal, including containers and MIG with Kubernetes . MIG is supported on GPUs starting with the NVIDIA Ampere generation models
https://docs.nvidia.com/datacenter/tesla/mig-user-guide/supported-gpus.html
Hardware Profile on A100‑40GB GPU
The raw hardware of a A100 GPU has:
MIG chops these 108 SMs into 7 compute slices and also the memory into 5 GB chunks
So using MIG profiles, we can create these combinations
So using MIG profiles, we can create these combinations as well..
Compute Instance
1 compute slice to create a 1g.5gb would like this
4g.5gb GI profile:
The diagram below shows a pictorial representation of how to build all valid combinations of GPU instances.
The assignments can be viewed by the nvidia-sm command
nvidia-smi -L
GPU 0: NVIDIA H100 80GB HBM3 (UUID: GPU-c08d91cb-e324-655c-71ba-7570956445bc)
MIG 1c.3g.40gb Device 0: (UUID: MIG-c788539f-c1ea-5a36-8b4d-9b07d024b1bb)
MIG 1c.3g.40gb Device 1: (UUID: MIG-405bbda1-6b05-535f-9702-f95e8cd170ce)
MIG 1c.3g.40gb Device 2: (UUID: MIG-df7e5a27-eeea-51a0-9055-b36d5a552879)
MIG 1g.10gb Device 3: (UUID: MIG-b4b2228d-6933-5839-bc0e-41ab9edb61c6)
MIG 1g.10gb Device 4: (UUID: MIG-c71dc464-b9f9-5611-9d29-d601a47cffd6)References
Here’s the full responsibility matrix for the VCF Admin role